Theses & Reports

Instructions for submitting a technical report or thesis.

You can find technical reports published prior to 1990 archived here.

• Ph.D. Thesis 2024 On Efficient Instantiations of Secure Multi-Party Computation in Practice Bienstock, Alexander Abstract | PDF

  Title: On Efficient Instantiations of Secure Multi-Party Computation in Practice

  Candidate: Bienstock, Alexander

  Advisor(s): Yevgeniy Dodis/Marshall Ball

  Abstract:

  Secure Multi-Party Computation (MPC) is an area of cryptography that has been studied extensively since the 1980s. In full generality, MPC allows a set of mutually distrusting parties to privately compute a function of their inputs. That is, the parties interact in some protocol, and at the end obtain the output of the function, and nothing else. In the decades since the inception of MPC, great strides have been made towards making it more efficient. However, despite this progress, the use of MPC in practice still faces some shortcomings.
  
  In this thesis, we take steps to mitigate two such shortcomings. The first deficiency we study is related to the communication networks in which such MPC protocols operate. MPC protocols are usually designed assuming that all parties have pairwise secure communication channels which are stable; i.e., nodes never crash, messages always arrive on time, etc. However, in the real-world, this is rarely the case—it is hard to sustain a stable connection between parties over long periods of time. One such model that has been introduced to address this deficiency is called Fluid MPC (Choudhuri et al., CRYPTO 2021). In this model, parties are not mandated to stay online for long periods of time. Instead, parties come online for short periods of time and work together in committees to compute some function. The benefit is that individual committees are much more likely to be able to sustain stable connections for these shorter interactions. However, existing protocols in this model do not match the level of efficiency that is obtained by traditional MPC protocols. In the first part of this thesis, we study Fluid MPC, and in particular, introduce Fluid MPC protocols with efficiency that matches those of traditional MPC.
  
  The second deficiency of MPC which we study in this thesis is that general-purpose protocols often are still not efficient enough to be used in practice. One way to resolve this is by using protocols that are tailor-made for specific applications. One such application that has gained recent attention is called Private Join and Compute (PJC). In this application, two parties come together with input sets and associated values for each item in their sets. The goal is to privately compute a function over the associated values of the intersection of the two sets. In practice, the size of the intersection is quite small, and therefore the private computation of the intersection is actually much more expensive than whatever computation that needs to be done over it. In the second part of this thesis, we improve the efficiency of tailor-made state-of-the-art protocols that are used to privately compute the intersection, thus improving the efficiency of prior PJC protocols.

• Ph.D. Thesis 2024 Generative modeling and Stochastic Control as Dynamics on Probability Distributions Domingo i Enrich, Carles Abstract | PDF

  Title: Generative modeling and Stochastic Control as Dynamics on Probability Distributions

  Candidate: Domingo i Enrich, Carles

  Advisor(s): Joan Bruna

  Abstract:

  Several modern machine learning algorithms can be studied from the perspective of evolution dynamics on the space of probability measures. Gradient descent-ascent algorithms that are used to solve minimax problems such as the ones arising in generative adversarial networks (GANs) can be interpreted as a joint evolution of two measures: one over the space of parameters of the generator, and one over the space of parameters of the discriminator.
  
  In the first chapter of the thesis, I study systems of this form, and provide convergence guarantees when possible. Diffusion models, which are another generative modeling technique, are also based on dynamics on probability measures, in this case over the space of samples. The dynamics are simulated at inference time; the starting distribution is a Gaussian, and the final distribution is meant to be the target data distribution. Diffusion models were generalized by the Flow Matching framework, which allows to construct different paths between the Gaussian noise distribution and the data distribution.
  
  In the second part, I introduce Multisample Flow Matching, which is a generalization of Flow Matching with intimate connections to optimal transport. Stochastic optimal control is a third problem where dynamics on
  measures play a critical role. The goal is to learn a vector field (the control) in order to drive the behavior of the solutions of a stochastic differential equation.
  
  In the third chapter, I present Stochastic Optimal Control Matching, a least-squares loss that is based on the same principles that are used to formulate diffusion model losses, and which achieves errors that are an order of magnitude lower than for existing methods.
  
  The talk will cover the second and third chapters.

• Ph.D. Thesis 2024 Solver-Aided Compiler Design for Programmable Network Devices Gao, Xiangyu Abstract | PDF

  Title: Solver-Aided Compiler Design for Programmable Network Devices

  Candidate: Gao, Xiangyu

  Advisor(s): Anirudh Sivaraman, Srinivas Narayana

  Abstract:

  Historically, network devices were mostly fixed-function ones. They could run at a line rate of one network packet per nanosecond, but it was impossible to support newly developed network algorithms without upgrading the device. The emergence of programmable network devices remedies this drawback. These devices use reconfigurable match table model to enable programmability and provide more flexibility for developers to continue updating and adding new algorithms to the device. People developed several programming languages to write programs for these devices. Even though it is not hard to get started with writing packet-processing code, writing programs that can fit within the target devices’ various resource constraints is not an easy job. The root cause for that is the lack of optimizing compilers in this domain. Hence, this thesis focuses on optimizing compiler design for domain-specific network accelerators using solver-aided techniques that can generate better compilation results compared with state-of-the-art compilers.

   

  First, we build the Chipmunk compiler that does code generation for stateful transactions into programmable switches using program synthesis. We frame the compilation problem as a solution-searching problem and use a program synthesis engine, SKETCH, to find a semantically equivalent compilation outcome. Additionally, we also develop a series of algorithms to speed up the compilation process. We find that the Chipmunk compiler can generate better compilation results in terms of hardware resource usage within a reasonable time period.

   

  Second, we build the CaT compiler that does both code generation and resource allocation into the packet-processing pipeline using solver-aided technologies. We decompose the compilation problem for such pipelines into three phases—making extensive use of solver engines to simplify the development of these phases. We also incorporate some heuristics for further resource usage optimization. We observe that the CaT compiler can generate near-optimal compilation results at a much faster speed than Chipmunk.

   

  Third, we build the Polyglotter compiler that outputs programs for target hardware devices from input programs written for source hardware devices in the parser portion. This compiler unifies features across different programming languages and reduces the efforts required to write algorithms across platforms. We discover that the Polyglotter compiler can generate correct transpilation results with better hardware resource usage.

   

  To our best knowledge, we for the first time propose to incorporate solver-aided techniques into compiler design for programmable network devices. In the domain of programmable network devices, these compilers can outperform traditional compilers that rely on program rewrite rules. Our contributions are beyond just building solver-aided compilers and include domain specific algorithms to speed up the whole compilation process. Based on these developed compilers, we explore several useful aspects of solver-aided techniques and hope to extend them into more applications in future works.

• Ph.D. Thesis 2024 Predictive and Generative Models of Protein Sequence and Structure Lin, Zeming Abstract | PDF

  Title: Predictive and Generative Models of Protein Sequence and Structure

  Candidate: Lin, Zeming

  Advisor(s): Yann LeCun

  Abstract:

  Historically, protein engineering has predominantly involved a bottom-up strategy, utilizing naturally occurring components as the building blocks. However, the problem of designing arbitrary protein sequences and structures for specific problems present significant challenges due to the complexity of biological systems. In this work, we tackle the problem of developing models of protein sequences and structures for prediction and generation. We show that neural networks can learn the patterns inherent to these systems and provide results for modeling protein through predicting protein structures from a given sequence and vice versa. Generative models can also model the unconditional distributions of protein sequence and
  structure.

  
  To model protein structures, we present an autoencoder architecture that can produce a wide array of protein backbones to model protein structures. These structures exhibit both local and global coherence in terms of secondary and tertiary structures. Using classical techniques to design sequences that fold to generated backbones, we show that the model can generate novel sequences which are validated in-silico. To generate better sequences for these backbones, we then present ESM-IF1, a model for fixed backbone protein design. We designed a large-scale system to predict millions of structures using AlphaFold. By training on the synthetic data, we were able to obtain state of the art results and obtain over 50% sequence recovery.

  
  We then scale large protein language models to 15 billion parameters (ESM-2) as an unconditional model of protein sequences. ESM-2 is capable of replacing multiple sequence alignment (MSA) features to obtain nearly state-of-the-art structure prediction results from a single sequence Removing MSA features gives a 60x speed up, allowing us to catalog the largest database of predicted protein structures. We open-sourced the ESM Metagenomic Atlas, a database of over 225 million high-confidence predicted structures, giving us an unprecedented view into the vast breadth and diversity of natural proteins. Finally, the speed and single sequence nature of our model allows us to directly optimize the protein sequence with respect to the protein structure. We show that black box optimization techniques can enable the design of proteins with structural constraints as symmetry, scaffolding, and binding. In sum, we present a series of models that are able to model the conditional and unconditional distributions of protein sequence and structure.

• Ph.D. Thesis 2024 Unlocking AI for Healthcare: Out-of-distribution Generalization and Causal Inference Manas Puli, Aahlad Abstract | PDF

  Title: Unlocking AI for Healthcare: Out-of-distribution Generalization and Causal Inference

  Candidate: Manas Puli, Aahlad

  Advisor(s): Prof. Rajesh Ranganath

  Abstract:

  Modern AI models make it easy to exploit the correlations in a dataset to predict a target of interest from a given set of inputs. However, the primary use of these models often lies outside the training data. For example, while one can train a Transformer to correlate a patient's medical history to their chances of developing coronary heart disease (CHD), the goal would be to estimate risks on populations elsewhere or in the future. Challenges arise if the model relies on correlations that shift between training and test times or capture non-causal relationships. Predictions based on unstable relationships can degrade outside the training distribution, and basing treatment decisions on non-causal relationships can result in harm. This thesis first develops a methodology for generalizing out-of-distribution (OOD) and estimating causal effects. It closes with an empirical study of building and transporting CHD risk models at two large hospital systems.
  
  The first part begins by defining a class of distribution shifts where standard training or balancing the data yield models can perform worse than random guessing. We characterize representations that generalize across such shifts and derive an algorithm to build models with such representations. Next, we develop an approach to encode knowledge of features used by humans into building robust models. The last work in this part identifies biases implicit in the standard way of training, gradient-based optimization of cross-entropy, that force models to depend more on unstable features than on the more informative stable ones. We develop a class of loss functions to encourage dependence on the more informative features.
  
  The second part of this thesis studies cases where common assumptions that enable causal estimation are violated. We provide an algorithm to estimate causal effects with deep models from confounded data where instrumental variables are available. This algorithm generalizes the control function method and works without the separability assumptions required by popular algorithms like the two-stage least-squares and generalized method of moments. Then, we consider tasks where the confounders are known to equal a function of the variables whose effects we want to estimate; this setup violates an assumption known as overlap or positivity, commonly made to uniquely determine (identify) causal effects from non-randomized data. In this setting, we derive nonparametric conditions for identifiability and derive an estimator that solves a gradient flow equation to answer general causal queries from the data without overlap.
  
  The last part of this thesis performs an empirical study of building and transporting CHD risk models between two large hospitals. Departing from the standard approach of constructing risk scores from carefully chosen features, we use broad feature sets available in the electronic health records (EHRs). We train AI models to predict time-to-CHD from minimally curated EHR data that outperform existing risk scores at the institution where they were trained and when transported externally.

• Ph.D. Thesis 2024 Towards Responsible AI: Safeguarding Privacy, Integrity, and Fairness Mirza, Muhammad Shujaat Abstract | PDF

  Title: Towards Responsible AI: Safeguarding Privacy, Integrity, and Fairness

  Candidate: Mirza, Muhammad Shujaat

  Advisor(s): Prof. Christina Pöpper

  Abstract:

  The widespread adoption of Artificial Intelligence (AI) into digital platforms, spanning general-purpose applications such as chatbots, professional tools like code generation, and high-risk domains like healthcare, has profoundly transformed user experiences. However, this rapid integration has also brought to the forefront critical concerns surrounding privacy, integrity, and fairness. This thesis systematically investigates these three interconnected challenges through comprehensive investigations revealing vulnerabilities and proposes approaches to address them, contributing to the responsible development of AI technologies.
  
  In addressing privacy concerns, we focus on managing personal information exposure in an era where digital data persists indefinitely. We begin with a global longitudinal analysis of privacy narratives to contextualize the evolving landscape of privacy concerns. Next, we systematically develop a semi-automated pipeline to assess the risks of training data extraction from large language models (LLMs), particularly those used for code generation such as Github Copilot. We demonstrate the feasibility of leaking various types of sensitive personal information, including email addresses, medical records, and passwords. Finally, we undertake a comprehensive systematization of privacy-enhancing technologies for exposure management, bridging gaps between technical solutions and user needs. We identify key discrepancies and propose actionable strategies for aligning technical solutions with user expectations. These findings lay the groundwork for user-centric privacy solutions that effectively address data persistence challenges.
  
  To tackle threats to information integrity, we focus on the potential misuse of generative AI tools and coordinated disinformation campaigns. We conduct a detailed evaluation of factual accuracy of frontier LLMs, such as the GPT series, in the zero-shot classification setting. By comparing different model versions we uncover inconsistencies in performance improvements, with GPT-4's March release outperforming its June counterpart. Next, we develop a novel cybersecurity-inspired framework for characterizing disinformation threats, profiling threat actors, attack patterns, targets, and channels. We validate our framework's effectiveness through case studies of real-world disinformation campaigns, highlighting its potential to strengthen the integrity of online information ecosystems and laying the groundwork for potential automated threat-scoring systems.
  
  Lastly, we address fairness in machine learning systems by identifying biases that reinforce inequalities. We introduce Global-Liar, a novel dataset uniquely balanced in terms of geographic representation, facilitating a more nuanced factuality evaluation of LLM biases across different regions. Using this dataset, we conduct a rigorous evaluation of general-purpose LLMs, revealing significant disadvantages faced by the Global South. Next, we conduct thorough investigation into fairness in high-risk computer vision models used for medical diagnosis in healthcare. Our assessment reveals significant racial and sex biases in kidney and tumor segmentation tasks. We investigate a range of bias mitigation approaches, from pre-processing techniques, like stratified batch sampling, to algorithmic interventions, like fair meta-learning. Notably, our findings suggest that architectural choices play a significant role in bias reduction, emphasizing the necessity of careful design and thorough evaluation of model architectures.
  
  In summary, our findings and proposed solutions in privacy, integrity, and fairness contribute to responsible AI development, aiming to democratize its benefits across all constituencies.

• Ph.D. Thesis 2024 Learning from Rewards in Text Generation Pang, Richard Yuanzhe Abstract | PDF

  Title: Learning from Rewards in Text Generation

  Candidate: Pang, Richard Yuanzhe

  Advisor(s): He He, Kyunghyun Cho

  Abstract:

  The progress in text generation comes from every stage in the pipeline: problem definition, data curation, learning, decoding, and evaluation. This dissertation focuses on learning. There is a mismatch between traditional training objectives and evaluation objectives: regular maximum likelihood estimation tries to minimize the cross-entropy loss with respect to each sample in the dataset, but the downstream evaluation is often based on a reward that scores the compatibility of the input-output pair (e.g., human judgments of the output). I aim to bridge this gap by optimizing for the reward of the generated text directly.
  
  The talk is composed of the following components. (1) Rewards could be expensive to obtain. To tackle this challenge in the social dialogue setting, we extract implicit signals from deployment data without extra human annotations. (2) The model could make slow or no progress in learning, and one idea is to obtain denser and high-quality rewards. In neural machine translation, we define a reward inspired by noisy channel decoding which has a long history, and we are able to increase decoding speed significantly while ensuring similar translation quality. (3) Another way to make progress in learning is to innovate on training algorithms instead. We set the rewards to be based on the simple exact match of generations and references, but algorithm-wise we explore the extreme case where we do not deviate too far from references by framing text generation as an offline reinforcement learning (RL) problem. We propose generation by off-policy learning from demonstrations (GOLD) using importance weighting. Our generations outperform those trained by MLE and policy gradient on a range of tasks. (4) We show that we do not need to rely on RL using a few reasoning tasks (e.g., math, science, commonsense) as the testbed. We develop an approach called iterative reasoning preference optimization (IRPO) that optimizes for winning vs. losing reasoning chain-of-thoughts, using modified direct preference optimization as the criteria. IRPO results in markedly increased accuracies compared to a range of baselines.
  
  To conclude the talk, I will discuss the future directions of using large language models as rewards. I will briefly mention the initial promise given by our work on self-rewarding language models using LLM-based rewards with a learning algorithm connected to that in IRPO; the discussion is then followed by the corresponding challenges and next steps. I will also touch on human–AI collaboration – an additional way to improve LLM evaluation capabilities.

• Ph.D. Thesis 2024 Verification of Concurrent Search Structures Patel, Nisarg Abstract | PDF

  Title: Verification of Concurrent Search Structures

  Candidate: Patel, Nisarg

  Advisor(s): Prof. Thomas Wies

  Abstract:

  Concurrent search structures are a class of concurrent data structures that implement a key-value store. Concurrent search structures are integral components of modern software systems, yet they are notoriously difficult to design and implement. In the context of concurrency, linearizability is the accepted notion of correctness of a data structure. Verifying linearizability of concurrent search structures remains a formidable challenge due to the inherent complexity of the underlying algorithms. So far, verification of these data structures has often led to large, intricate proofs that are hard to comprehend and reuse.
  
  The concrete contribution of the thesis is developing and verifying new template algorithms that cover several variants of lock-free skiplists and lock-based log-structured merge (LSM) trees. The template algorithms capture concurrency mechanism, but abstract away node-level details and the maintenance operations.
  
  The generalizable contribution of the thesis is the advancement in the verification technology required to prove the new template algorithms. There are two key contributions here, first relating to hindsight reasoning and second to keyset reasoning. Hindsight reasoning has been shown to be useful for proving linearizability, but it has not been explored in the context of a foundational program logic. The thesis addresses the challenge by embedding the technique of hindsight reasoning in the concurrent separation logic Iris via prophecy variables. Keyset reasoning is useful for lifting assertions on a node's contents to the global contents held by the structure. The thesis develops a keyset resource algebra, an Iris resource algebra to enable keyset reasoning in Iris.
  
  All of the techniques and proofs are mechanized in Iris/Coq. Verified search structures include in particular the Michael set, the Harris list, the Herlihy-Shavit skiplist and an LSM-tree implementation based on LevelDB. The verification effort represents a significant contribution as it is the first mechanized proof of linearizability for concurrent skiplists and LSM-trees.

• Ph.D. Thesis 2024 Neural Language Representations and Scaling Semi-Supervised Learning for Speech Recognition Peyser, Cal Abstract | PDF

  Title: Neural Language Representations and Scaling Semi-Supervised Learning for Speech Recognition

  Candidate: Peyser, Cal

  Advisor(s): Prof. Kyunghyun Cho, Prof. Michael Picheny

  Abstract:

  Speech recognition research has been focused for several years on the incorporation of unpaired speech and text data alongside conventional supervised datasets. Dominant methods have emphasized auxiliary tasks for refining speech and/or text representations during model training. These methods have generally performed strongly when paired with very small supervised datasets, but do not yield the same improvements against strong, supervised baselines. We argue in this thesis that the path to scaling these methods lies in the speech and text representations themselves. We investigate statistical properties of these representations, and show that downstream ASR performance corresponds to a model's ability to jointly represent speech and text. We analyze existing methods for semisupervised ASR, and develop an algorithm to improve them at scale by aligning speech and text in representation space.

• Ph.D. Thesis 2024 DrawTalking: Building Interactive Worlds by Sketching and Speaking Rosenberg, Karl Toby Abstract | PDF

  Title: DrawTalking: Building Interactive Worlds by Sketching and Speaking

  Candidate: Rosenberg, Karl Toby

  Advisor(s): Ken Perlin

  Abstract:

  This thesis introduces the design and implementation of an interaction concept called DrawTalking. Through simple combinations of sketching and speaking, the user can improvisationally build an interactive world of graphics, animations, diagrams, and dynamic mechanisms with behavior and rules, as if by narrating a story or explaining a concept to an audience. The interface demonstrates a possible step towards designing future interfaces more closely in-tune with how we naturally communicate and think.
  
  For context, sketching while speaking has played a major part in innovation across disciplines. The combination of visuals and spoken language enables us to make-believe: think about, describe, communicate, and interact with anything that we can think of, including things that do not or cannot exist in the real world. Evolving technology creates opportunities to move beyond sketching and speech alone. Human-computer interactions of the future, drawing inspiration from our process of make-believe, can add interactive computation to the combination of sketching and speech, allowing us to work with explorable worlds, simulations, and mechanics. By enabling such interactions, we might think, learn, design, play, and tell stories in increasingly expressive ways.
  
  Towards this idea, what makes for a good interface for computation-mediated sketching and speaking? This touches upon several fundamental questions in interaction design, human-AI interaction, and human-centered interfaces, chiefly among them, how to balance human control and machine automation?
  
  Inspired by real-world speaking and sketching interactions, and seminal works in dynamic sketching, interactive visual programming, and language interfaces, we designed interaction techniques that draw on the way people describe objects and phenomena when telling stories and explaining processes at a whiteboard.
  
  How does it work? the user speaks to label hand-drawn sketches with names and properties, and to define rules for how their world should behave. This communicates semantic intent to the computer, while giving the user the flexibility to choose how to represent and change their drawings. Now the user can interact with a simulated world simply by narrating stories or describing mechanics, which dynamically creates running interactive programs from built-in primitives and user-customized rules.
  
  To gauge understanding of the mechanics of DrawTalking and to derive use cases, we invited participants to an open-ended one-on-one user-study session with the researcher to discover and explore the features in DrawTalking. Each user improvised and prototyped interactive sketch-based animations and gameplay scenarios by collaborating with the researcher. The resulting artifacts and discussion were oriented around each participant's specific experiences and background.
  
  Feedback suggests that our approach is promising and intuitive: it prioritizes user control; it is flexible and supports improvisation; the workflow is fluid; the features are extensible and adaptable to other application domains and contexts beyond sketching; the design demonstrates how multiple applications can use similar language-based interaction techniques and behaviors predictably alongside other language-based technologies; it enables programming-like capability without code.
  
  Through the research and design process of DrawTalking, we learned that it could represent an approach to designing complex interoperating systems for human-AI collaboration. We hope it can serve as a useful example for research and design of future machine-mediated interfaces, interactions, and computer systems.

• Ph.D. Thesis 2024 Algorithmic enhancements to causal inference problems Shen, Bingran Abstract | PDF

  Title: Algorithmic enhancements to causal inference problems

  Candidate: Shen, Bingran

  Advisor(s): Prof. Dennis Shasha

  Abstract:

  This thesis explores novel approaches to inferring and representing causal relationships in biological networks. We introduce EnsInfer, an ensemble method that combines state-of-the-art inference algorithms using a Naive Bayes classifier, outperforming individual methods and providing a flexible framework for integrating diverse data types. Our research then challenges the conventional representation of gene regulatory networks (GRNs) by demonstrating that nonlinear machine learning models achieve better predictive performance than models based solely on "gold standard" regulatory edges. To address this limitation, we propose a bipartite network representation that better captures the synergistic regulatory effects of multiple transcription factors on target genes. This framework focuses on four key goals: predictive accuracy, parsimonious enumeration of predictive regulatory genes, identification of disjoint sets of predictive regulatory genes, and construction of a bipartite network representation of causality. Our work provides an actionable and interpretable paradigm for investigating causal gene regulation, with potential applications across diverse domains of causality research.

• Ph.D. Thesis 2024 Olympiad-level Geometry Theorem Proving without Human Demonstrations Trinh, Trieu Abstract | PDF

  Title: Olympiad-level Geometry Theorem Proving without Human Demonstrations

  Candidate: Trinh, Trieu

  Advisor(s): He He

  Abstract:

  Proving mathematical theorems at Olympiad level represents a significant milestone in human-level automated reasoning, owing to their reputed difficulty among the world’s best talents in pre-university mathematics. Current machine learning approaches, however, are not applicable to most mathematical domains due to the high cost of translating human proofs into machine-verifiable format. The problem is even worse for geometry due to its unique translation challenges, resulting in severe scarcity of training data. We propose G0, a theorem prover for Euclidean plane geometry that sidesteps the need for human demonstrations by synthesizing millions of theorems and proofs across different levels of complexity. G0 is a neuro-symbolic system that uses a neural language model, trained from scratch on our large-scale synthetic data, to guide a symbolic deduction engine through infinite branching points in challenging problems. On a test set of 30 latest Olympiad problems, G0 solves 25, outperforming the previous best method that only solves 10 problems and approaching the performance of an average International Mathematical Olympiad (IMO) gold medalist. Notably, G0 produces human-readable proofs, solves all geometry problems in the IMO 2000 and 2015 under human expert evaluation, and discovers a generalized version of a translated IMO theorem in 2004.

• Ph.D. Thesis 2024 Improve Language Model Serving Efficiency with Fine-grained and Stateful Scheduling Yu, Lingfan Abstract | PDF

  Title: Improve Language Model Serving Efficiency with Fine-grained and Stateful Scheduling

  Candidate: Yu, Lingfan

  Advisor(s): Jinyang Li

  Abstract:

  The world has witnessed the remarkable success of large language models (LLMs), led by the fast-growing popularity of ChatGPT. However, it is challenging to serve these language models and deliver both high throughput and low latency due to the iterative nature of language models. This thesis identifies two key issues impacting the performance of existing systems: (1) Coarse-grained batching at the request level results in wasteful computation for requests with variable input and output lengths; (2) The lack of stateful context management results in duplicate computation for applications that engage in multi-turn interactions with the LLM model. Two systems, BatchMaker and Pensieve, are then presented to address these issues.
  
  BatchMaker proposes a technique called cellular batching to improve the latency and throughput of language model inference. Existing systems use batch execution of the dataflow graphs of a fixed set of requests. By contrast, BatchMaker makes finer-grained batching decisions at each token processing step, and dynamically assembles a batch for execution as requests join and leave the system.
  
  Pensieve is a system optimized for multi-turn conversation LLM serving. It maintains the conversation state across requests from the same conversation by caching previously processed history to avoid duplicate processing. Pensieve's multi-tier caching strategy utilizes both GPU and CPU memory to store and retrieve cached data efficiently. Pensieve also generalizes the recent PagedAttention kernel to support attention between multiple input tokens whose KV cache is spread over non-contiguous GPU memory.
  
  Experiments on various workloads show that BatchMaker improves throughput by 25-80% while reducing latency by 18-90% latency, and Pensieve improves throughput by 33-100% and reduces latency by 40-77%.

• Ph.D. Thesis 2024 Theory of Symmetric Neural Networks Zweig, Aaron Abstract | PDF

  Title: Theory of Symmetric Neural Networks

  Candidate: Zweig, Aaron

  Advisor(s): Joan Bruna

  Abstract:

  Symmetric functions, which take as input an unordered, fixed-size set, find practical application in myriad physical settings based on indistinguishable points or particles, and are also used as intermediate building blocks to construct networks with other invariances. Symmetric functions are known to be universally representable by neural networks that enforce permutation invariance. However the theoretical tools that characterize the approximation, optimization and generalization of typical networks fail to adequately characterize architectures that enforce invariance.
  
  This thesis explores when these tools can be adapted to symmetric architectures, and when the invariance properties lead to new theoretical findings altogether. We study and prove approximation limitations on the extension of symmetric neural networks to infinite-sized inputs, the approximation capabilities of symmetric and antisymmetric networks relative to the interaction between set elements, and the learnability of simple symmetric functions with gradient methods

• Ph.D. Thesis 2023 On Matching Problems in Large Settings Agarwal, Ishan Abstract | PDF

  Title: On Matching Problems in Large Settings

  Candidate: Agarwal, Ishan

  Advisor(s): Richard Cole

  Abstract:

  Matching problems arise in several settings in practice and have been a longstanding subject of theoretical analysis. Typically, the settings of interest involve a large number of agents. We further the study of matching problems in two settings: the stable matching setting, which has been studied since the seminal work of Gale and Shapley, and a setting where agents' values to prospective partners degrade over time, leading them to have to balance the trade-off between searching for a better partner versus deciding to match.
  
  In the stable matching setting, we extend a line of research that seeks to explain the dichotomy between the fact that Gale and Shapley's Deferred Acceptance algorithm seems to work well in practice, even when agents only submit a short list of prospective partners to the centralized matching algorithm, and the fact that if the agents' preferences are allowed to be arbitrary, complete lists of all agents' preferences are needed in order to guarantee a stable matching. To this end, we consider probabilistically generated preference lists and we show that under fairly general assumptions and in a variety of models, with high probability, short lists of prospective partners, namely length $\Theta (\log n)$ instead of $n$, suffice for most of the agents. We prove our bounds are tight up to constant factors. Furthermore, we construct a simple set of $\Theta (\log n)$ possible matches per agent for almost all agents and demonstrate (in the form of an approximate equilibrium result) that they can afford to restrict their proposals to this set, while incurring only a small loss in utility.
  
  In the time discounted utilities setting, we consider a dynamic matching market, and study how agents should balance accepting a proposed match with the cost of continuing their search. Our model has two new features: finite agent lifetimes with linear loss in utility over time, and a discrete population model, aspects which are underexplored in the literature. We quantify how well the agents can do by providing upper and lower bounds on the collective losses of the agents, with a polynomially small failure probability, where the notion of loss is with respect to a plausible baseline we define. These bounds are also tight up to constant factors.
  
  In both settings, we complement our theoretical results with numerical simulations.
   

• Ph.D. Thesis 2023 Function Space Reasoning in Gaussian Processes and Neural Networks Benton, Gregory Abstract | PDF

  Title: Function Space Reasoning in Gaussian Processes and Neural Networks

  Candidate: Benton, Gregory

  Advisor(s): Andrew Gordon Wilson

  Abstract:

  In a typical modeling setting we have prior notions of what types of functions we want to learn. For example, in regression we may want to learn a smooth function or a periodic function and in image classification we may want to learn a function that is invariant to rotations. While function space provides us the benefit of being able to reason about traits like invariance or smoothness, it is often difficult to directly quantify the functional properties of models, in particular for large parametric models like neural networks.
  
  In this thesis we leverage our ability to reason about function space to build more powerful models in both Gaussian processes (GPs) and neural networks. By generating GP kernels as functions themselves of latent processes, we introduce methods for providing uncertainty over what types of functions we produce, not just over the functions themselves in GP models. We also introduce methods for learning levels of invariance and equivariance in neural networks, enabling us to imbue the functions our models produce with soft or limited equivariance constraints. Finally, we show how we can leverage our understanding of parameter space in neural networks to efficiently ensemble diverse collections of functions to improve the accuracy and robustness of our models. Through the introduction of these methods we show that by carefully considering the types of functions we are producing we can describe models with a range of desirable properties. These properties include more flexible models, models that better align with domain knowledge, and models that are both accurate and robust. We demonstrate these results on a broad range of problems, including time series forecasting, image classification, and reinforcement learning.
   

• Ph.D. Thesis 2023 Bridging the Gap from Supervised Learning to Control Brandfonbrener, David Abstract | PDF

  Title: Bridging the Gap from Supervised Learning to Control

  Candidate: Brandfonbrener, David

  Advisor(s): Joan Bruna

  Abstract:

  he combination of deep learning and internet-scale data with supervised
  learning has led to impressive progress in recent years. However, the
  potential of this progress has yet to be realized in the context of
  control problems beyond games that are easy to simulate. This thesis
  attempts to bridge this gap so as to leverage tools from supervised
  learning to solve control problems. To do this, we focus on the offline
  reinforcement learning setting which attempts to learn a control policy
  from a fixed dataset rather than requiring the policy to learn and
  collect data at the same time. This removes issues of non-stationary
  training data and exploration from the control problem, which allows the
  more straightforward application of tools from supervised learning.
  
  We study this intersection between supervised learning and control from
  several angles. In the first part of the thesis, we present work on
  policy learning, focusing on simplified algorithms that look more like
  standard supervised algorithms. In the second part, we move one step
  earlier in the pipeline and consider how to best collect datasets for
  offline reinforcement learning. And in the last part, we consider how to
  design pretraining objectives to learn representations for downstream
  offline policy learning. Taken together, these contributions present a
  view of the promise and challenges that face the application of machine
  learning to control problems.
   

• Ph.D. Thesis 2023 Provably Robust and Accurate Methods for Rigid and Deformable Simulation with Contact Ferguson, Zachary Abstract | PDF

  Title: Provably Robust and Accurate Methods for Rigid and Deformable Simulation with Contact

  Candidate: Ferguson, Zachary

  Advisor(s): Daniele Panozzo

  Abstract:

  Contacts are essential to virtually every aspect of life and play a vital role in many physical phenomena. Because of this, the study of contact mechanics has a deep wealth of knowledge. Surprisingly, however, simulating contact is a challenge with many parameters to carefully adjust. Incorrect parameters can result in numerical explosions, intersections, and other failures. Our research seeks to address these problems by developing robust methods that can handle arbitrary scenarios with guaranteed success.
  
  In this thesis, we introduce the Incremental Potential Contact (IPC) method. IPC is the first simulation algorithm for deformable and rigid bodies that is unconditionally robust, requires minimal parameter tuning, and provides a direct way of controlling the trade-off between running time and accuracy. We further back up these claims by providing a large-scale benchmark of continuous collision detection (CCD) algorithms (a core component of the IPC method) based on their efficiency and correctness. As part of this study, we introduce the first efficient CCD algorithm that is provably conservative. For extended accuracy and efficiency, we show how nonlinear geometry and function spaces can be used within the IPC framework. Finally, we introduce the first physically-based adaptive meshing strategy which produces more accurate discretizations depending on elastic, contact, and frictional forces.
  
  This work and our open-source implementations have quickly garnered attention from the computer graphics, mechanical engineering, and biomechanical engineering communities for their robustness and ability to seamlessly handle scenarios that have long been a challenge. This marks a large step towards democratizing simulation tools for design, robotics, biomechanical, and visual effects applications, among others.
  
   

• Ph.D. Thesis 2023 Understanding and Incorporating Mathematical Inductive Biases in Neural Networks Finzi, Marc Abstract | PDF

  Title: Understanding and Incorporating Mathematical Inductive Biases in Neural Networks

  Candidate: Finzi, Marc

  Advisor(s): Andrew Gordon Wilson

  Abstract:

  To overcome the enormous sample complexity of deep learning models, we can leverage basic elements of human and scientific knowledge and imbue these elements into our models. By doing so, we can short-circuit the thousands of years of evolutionary development that has enabled such rapid learning in humans, and the development of science which provides a framework to fit new knowledge into. In this work I develop new methods for incorporating mathematical inductive biases into our models, biasing them towards solutions that reflect our priors and our knowledge. This work helps to broaden the scope and automation of equivariant model construction across diverse domains, uncover the role of inductive biases in learning and generalization, and developing new machine learning models for scientific applications, capturing relevant scientific knowledge.

• Ph.D. Thesis 2023 Deconstructing Models and Methods in Deep Learning Izmailov, Pavel Abstract | PDF

  Title: Deconstructing Models and Methods in Deep Learning

  Candidate: Izmailov, Pavel

  Advisor(s): Andrew Gordon Wilson

  Abstract:

  Machine learning models are ultimately used to make
  decisions in the real world, where mistakes can be incredibly costly.
  We still understand surprisingly little about neural networks and the
  procedures that we use to train them, and, as a result, our models are
  brittle, often rely on spurious features, and generalize poorly under
  minor distribution shifts. Moreover, these models are often unable to
  faithfully represent uncertainty in their predictions, further
  limiting their applicability. In this dissertation, I present results
  on neural network loss surfaces, probabilistic deep learning,
  uncertainty estimation and robustness to distribution shifts. In each
  of these works, we aim to build foundational understanding of models,
  training procedures, and their limitations, and then use this
  understanding to develop practically impactful, interpretable, robust
  and broadly applicable methods and models.

• Ph.D. Thesis 2023 Learning structured and stable reduced models from data with operator inference Sawant, Nihar Abstract | PDF

  Title: Learning structured and stable reduced models from data with operator inference

  Candidate: Sawant, Nihar

  Advisor(s): Benjamin Peherstorfer

  Abstract:

  Operator inference learns low-dimensional dynamical-system models with polynomial nonlinear terms from trajectories of high-dimensional physical systems (non-intrusive model
  reduction). This work focuses on the large class of physical systems that can be well described by models with quadratic and cubic nonlinear terms and proposes a regularizer for
  operator inference that induces a stability bias onto learned models. The proposed regularizer is physics informed in the sense that it penalizes higher-order terms with large norms and
  so explicitly leverages the polynomial model form that is given by the underlying physics.
  This means that the proposed approach judiciously learns from data and physical insights
  combined, rather than from either data or physics alone. A formulation of operator inference
  is proposed that enforces model constraints for preserving structure such as symmetry and
  definiteness in linear terms. Additionally, for a system of nonlinear conservation laws, we
  enforce model constraints that preserve the entropy stability of the dynamical system. Numerical results demonstrate that models learned with operator inference and the proposed
  regularizer and structure preservation are accurate and stable even in cases where using no
  regularization and Tikhonov regularization leads to models that are unstable.

• Ph.D. Thesis 2023 Continuous LWE and its Applications Song, Min Jae Abstract | PDF

  Title: Continuous LWE and its Applications

  Candidate: Song, Min Jae

  Advisor(s): Oded Regev/Joan Bruna

  Abstract:

  Efficiently extracting useful information from high-dimensional data is a major challenge in machine learning (ML). Oftentimes, the challenge comes not from a lack of data, but from its high dimensionality and computational constraints. For instance, when data exhibits a low-dimensional structure, one could in principle exhaustively search over all candidate structures, and obtain estimators with strong statistical guarantees. Of course, such brute-force approach is prohibitively expensive in high dimensions, necessitating the need for computationally efficient alternatives. When our problem, however, *persistently* eludes efficient algorithms, we may find ourselves asking the following perplexing question: is the failure due to our lack of algorithmic ingenuity or is the problem just too hard? Is there a *gap* between what we can achieve statistically and what we can achieve computationally?
  
  This thesis is one attempt at answering such questions on the computational complexity of statistical inference. We provide results of both positive and negative nature on the complexity of canonical learning problems by establishing connections between ML and lattice-based cryptography. The Continuous Learning with Errors (CLWE) problem, which can be seen as a continuous variant of the well-known Learning with Errors (LWE) problem from lattice-based cryptography, lies at the center of this fruitful connection.
  
  In the first part of this thesis, we show that CLWE enjoys essentially the same average-case hardness guarantees as LWE. This result has several important applications. For example, it shows that estimating the density of high-dimensional Gaussian mixtures is computationally hard, and gives rise to "backdoored" Gaussian distributions that can be used to plant undetectable backdoors in ML models and construct novel public-key encryption schemes.
  
  Next, we focus on the "backdoored" Gaussian distributions, which we refer to as Gaussian Pancakes, and the problem of distinguishing these distributions from the standard Gaussian. We provide several evidence for the hardness of this distinguishing problem based on a reduction from CLWE and lower bounds against restricted classes of algorithms, such as algorithms that compute low-degree polynomials of the observations.
  
  Finally, we end on a positive note by showing that the Lenstra-Lenstra-Lovasz (LLL) algorithm, commonly used in computational number theory and lattice-based cryptography, has surprising implications for noiseless inference. In particular, we show that LLL solves both CLWE and Gaussian Pancakes in the noiseless setting, in spite of the low-degree lower bound for Gaussian Pancakes. Furthermore, we show that LLL surpasses Sum-of-Squares and Approximate Message Passing algorithms, two methods often conjectured to be optimal among polynomial-time algorithms, on other noiseless problems such as Gaussian Clustering and Gaussian Phase Retrieval. These results highlight the crucial but subtle role of noise and hidden algebraic structure in the onset of statistical-to-computational gaps.

• Ph.D. Thesis 2023 Expanding Structural Design through Shape Optimization and Microstructures Tozoni, Davi Colli Abstract | PDF

  Title: Expanding Structural Design through Shape Optimization and Microstructures

  Candidate: Tozoni, Davi Colli

  Advisor(s): Denis Zorin

  Abstract:

  3D printing and other modern manufacturing tools allow users to design and produce customized objects for their needs at a considerably low cost. However, designing structures that are able to perform well is not an easy task and doing it manually can be a very slow and tedious process. In this context, structural optimization techniques can be very useful and help automating the design and analysis process.
  
  This thesis describes techniques that can expand the usage of structural optimization for digital fabrication by formulating optimization to be used with simulation models that are closer to reality, through the addition of contact and friction. Moreover, we show a fast method to compute gradients from differentiable simulations, which can be used to optimize shape, material and physical properties of our domain. In addition, we provide ways of expanding the use of two-scale topology optimization by presenting microstructures that have a smooth map from material to geometry and which can be used on curved shapes defined by irregular lattices with close to rhombic cells. Finally, we introduce two low-parametric microstructures that together are able to cover almost the whole possible range of elastic properties for isotropic metamaterials.
  
  Our results in simulation and physical experiments, both for static and time-dependent scenarios, show the advantages of our techniques and how they can be used in practice.

• Ph.D. Thesis 2022 Enhancing Robustness through Domain Faithful Deep Learning Systems Balashankar, Ananth Abstract | PDF

  Title: Enhancing Robustness through Domain Faithful Deep Learning Systems

  Candidate: Balashankar, Ananth

  Advisor(s): Lakshminarayanan Subramanian

  Abstract:

  In high-stakes domains like health, socio-economic inference, and content moderation, a fundamental roadblock for relying on deep learning systems is that models' predictions diverge from established domain knowledge when deployed in the real world and fail to faithfully incorporate domain-specific structure. In this talk, I will focus on the design of Domain Faithful Deep Learning Systems, that translate expert-understandable domain knowledge and constraints to be faithfully incorporated into learning robust deep learning models. Through methodological contributions in causal-aware ML model design, constrained optimization, counterfactual data augmentation, and feature selection, I have addressed core research questions of “What data distributions do domain practitioners care about?'', “How to faithfully convert domain knowledge into model constraints for better generalization?'' and finally ``How to evaluate whether the ML models we learn are grounded in the domain knowledge and in what ways do they deviate?''. I will demonstrate how, through these new approaches to incorporating domain knowledge, I have been able to meaningfully improve performance in four real-world applications of news-based famine forecasting, medication recommendations, causal question answering, and toxicity detection in online social media. These causal-aware and robust prediction models I have developed in collaboration with the World Bank and Google have shown that incorporating domain-specific structure is essential for building robust predictive models.

• Ph.D. Thesis 2022 Unstructured Mesh Generation and Repairing in the Wild Hu, Yixin Abstract | PDF

  Title: Unstructured Mesh Generation and Repairing in the Wild

  Candidate: Hu, Yixin

  Advisor(s): Daniele Panozzo

  Abstract:

  A mesh is a representation used to digitally represent the boundary or volume of an object for manipulation and analysis. Meshes can be used in many fields, including physical simulation in manufacturing, architecture design, medical scan analysis. In this thesis, we propose a series of meshing algorithms, named WildMeshing, that tackles one of the long-standing, yet fundamental, problems in geometry modeling: robustly and automatically generating high-quality triangle and tetrahedral meshes and repairing imperfect geometries in the wild. Different from existing methods that have assumptions about the input and thus often fail on real-world input geometries, WildMeshing provides strict guarantees of termination and is a black box that can be easily integrated into any geometry processing pipelines in research or industry.

  This thesis first investigates the problem of tetrahedralizing 3D geometries represented by piecewise linear surfaces. We propose an algorithm, TetWild, that is unconditionally robust, requires no user interaction, and can directly convert a triangle soup into an analysis-ready volumetric tetrahedral mesh. It relies on three core principles: hybrid geometric kernel, tolerance of the mesh relative to the surface input, and iterative mesh optimization with guarantees on the output validity. We then consider improving the algorithm efficiency for tetrahedralizing large-scale geometries. We design a new algorithm, fTetWild, that is based on the principles of TetWild but replaces the hybrid kernel with a floating-point kernel, which largely reduces runtime while keeping the same robustness. Next, this thesis explores meshing curved geometries. We start from the problem of triangulating 2D planar shapes whose boundaries are represented by curves. We introduce TriWild, an algorithm to robustly generate curved triangle meshes reproducing smooth feature curves, which leads to coarse meshes designed to match the simulation requirements necessary by applications and avoids the geometrical errors introduced by linear meshes.

  We test our algorithms on over ten thousand real-world input geometries and they achieve 100% success rate. Our methods generate meshes without any assumptions about the input while repairing the imperfect geometries, opening the door to automatic, large-scale processing of real-world geometric data.

• Ph.D. Thesis 2022 Data-driven Solutions for Addressing Two Pressing Urban Sustainability Challenges: Air Pollution Reduction and Traffic Management Iyer, Shiva Abstract | PDF

  Title: Data-driven Solutions for Addressing Two Pressing Urban Sustainability Challenges: Air Pollution Reduction and Traffic Management

  Candidate: Iyer, Shiva

  Advisor(s): Lakshmi Subramanian

  Abstract:

  Data Science and AI-driven solutions are abundant today for a large variety of practical applications. With a continuing focus on urban development and sustainability, in this thesis, I present our attempts in addressing two prominent urban challenges – urban air pollution control and road traffic congestion management. For both these applications, we have developed novel methods, such as the message-passing recurrent neural network, for predictive analytics and inference. The city of Delhi has 32 air quality monitors over an area of about 900 sq km, but we do not have information on fine-grained variations in air quality in the city in order to reason about citizen exposure and identify hotspots. We have installed 28 low-cost sensors, many of them concentrated in the south Delhi region. We have developed a generic definition of "hotspots" in terms of spatio-temporal variations, using which we validate some known hotspots and discover new ones. We have also designed a novel model combining geostatistics and deep learning that is able to make spatio-temporal pollution predictions by the hour with an MAPE of about 10% across all locations.

  In the context of urban traffic management, we first show that road networks can experience traffic jams over prolonged periods such as several hours due to sudden traffic bursts over short time scales. We illustrate this using real data from two different cities – New York and Nairobi. We provide a formalism for understanding the phenomena of traffic collapse and sudden jams. In the second work, we devise a novel model called the message-passing neural network for modeling the propagation of congestion within a road network and forecasting congestion. The MPRNN achieves the lowest mean error of < 0.3 mph when predicting ahead in 10 minute intervals, for up to 3 road segments ahead (message passing across 3 hops). Finally, in the third work, we describe an algorithm for signal control in free-flow road networks, inspired from congestion control in computer networks. Our proposed method significantly enhances the operational capacity of free-flow road networks in the real world by several orders of magnitude (between 3× and 5×) and prevents congestion collapse.

• Ph.D. Thesis 2022 Synergistic Geometry Processing: from Robust Geometric Modeling to Scalable Physical Simulation Jiang, Zhongshi Abstract | PDF

  Title: Synergistic Geometry Processing: from Robust Geometric Modeling to Scalable Physical Simulation

  Candidate: Jiang, Zhongshi

  Advisor(s): Daniele Panozzo

  Abstract:

  Various applications, from artistic creation, to scientific computing,
  require the processing and reasoning of 3D digital objects.
  The computational modeling of 3D geometric shapes, materials, and
  textures, as well as the simulation of their deformation and
  interactions, is essential to bring the algorithmic power of computing
  to real-life manufacture, architecture, and medical device design.
  Depending on the specific numerical properties, better algorithm
  designs might prefer 3D data with different representations, for
  example, in planes, surfaces, or inside volumes.

  This thesis investigates the problem related to the representations of
  data on 3D shapes and across different domains,
  so computations for different stages within a pipeline, may come
  together synergistically without manual tuning that disrupts an
  automated data flow.
  
  I propose novel geometrical principles in various geometric modeling
  and processing stages.  I also showcase various geometric computing
  applications that easily integrate such principles to guarantee the
  geometry validity and algorithm effectiveness of surface
  parameterization, rendering, deformation/animation, and mechanical
  simulation.
  In addition, we can finally explore creative solutions that reliably
  coarsen the surface. Such simplification accelerates everyday
  geometric modeling operations; the contribution also includes a
  scalable method to construct coarse and curved meshes for fast
  animation and scientific computing.
  
  Furthermore, the thesis provides a declarative way to formulate mesh
  processing and adaptation algorithms to facilitate the practical
  development of robust and reliable mesh processing software.
  Finally, the thesis includes extensive numerical validations involving
  tens of thousands of complex geometry shapes. And to maintain
  replicability and foster further research in this direction, I also
  released the implementation and generated data to be open source and
  accessible.
  
  Finally, the thesis includes extensive numerical validations involving
  tens of thousands of complex geometry shapes. %And to maintain
  replicability and foster further research in this direction, I also
  released the implementation and generated data to be open source and
  accessible.
   

• Ph.D. Thesis 2022 Cryptography: From Practice to Theory Karthikeyan, Harish Abstract | PDF

  Title: Cryptography: From Practice to Theory

  Candidate: Karthikeyan, Harish

  Advisor(s): Yevgeniy Dodis

  Abstract:

  This work is yet another attempt to turn an age-old adage on its head by deriving inspiration for theoretical research from problems that are germane to practitioners and real-world deployment. This could be viewed as a departure from the practice of creating real-world solutions that trace their origin to a theoretical research, or alternatively ex post facto theoretical analyses of practically deployed solutions that can be rather ad-hoc. Specifically, we look at four different problems that are relevant for practical deployment - random number generation, provably secure block ciphers, searching over encrypted data, and forward-secure group messaging.

• Ph.D. Thesis 2022 Scalable Distributed Payment Systems with Minimal Trust Assumptions Kattis, Assimakis Abstract | PDF

  Title: Scalable Distributed Payment Systems with Minimal Trust Assumptions

  Candidate: Kattis, Assimakis

  Advisor(s): Prof. Joseph Bonneau

  Abstract:

  Over the last decade, the security and resilience of Bitcoin
  as a stable payment network has motivated substantial study of the
  viability of distributed payment protocols, with many works focusing on their suitability as alternatives to centralized payment processing. We investigate the design of scalable distributed payment systems in the permissionless setting, where no actors in the protocol can be trusted or identified with out-of-band information. Scalability is identified with two desirable properties: high transaction processing rate (or throughput) and low confirmation latency (or settlement times). We analyze the trade-offs inherent to distributed protocols that prevent
  naive optimization of the above parameters and study techniques from verifiable computation as potential tools for overcoming these
  bottlenecks.
  
  One technique to increase throughput in distributed payment systems
  involves the use of Succinct Non-interactive ARguments of Knowledge
  (SNARKs, or SNARK proofs) to verify the integrity of transactions.
  Transaction rollups are one such solution, using SNARK computations to achieve scalability. Many instantiations of rollups leveraging SNARKs show encouraging evidence that this technique could achieve commercial- capacity throughput rates if implemented on top of current distributed payment systems, even in the smart-contract setting. Although promising, all rollup approaches require the resolution of an additional yet crucial question. For protocols operating in the permissionless setting, we need to ensure that a system relying on proof generation to scale also incentivizes actors to compute proofs cheaply and quickly. This is a governance problem, as the protocol needs to decide on how participants will be chosen to perform these (expensive) computations. We pose the question of who will compute the proofs, identify it as a consensus problem and provide a technical proposal towards its resolution.
  
  Our main contributions are twofold: in Part I, we design a
  permissionless consensus protocol that solves the problem of state
  verification for resource-limited clients in an incentive-compatible way. We show formal proofs of security and achieve minimal resource requirements for full ledger verification. This protocol showcases our key contribution: the design of a proof-of-work (PoW) process that computes SNARK proofs as valid outputs. Suitably choosing the statement whose proof is generated through PoW provides an incentive-compatible way to enforce the computation required by proof-based scaling techniques. In Part II, we look at one of the key components of SNARK- based throughput optimization: the non-interactive proof itself. We design a novel proof system which provides security guarantees in the trustless setting, while still being small and efficiently computable.
  This proof system (a transparent SNARK, or STARK) can be used directly for scaling throughput in distributed payments through transaction rollups. In conjunction with an incentivized PoW process, it also demonstrates a way for participants in consensus to quickly generate the rollup proofs in a permissionless way.
   

• Ph.D. Thesis 2022 Characterizing and Resolving Degeneracies in Neural Autoregressive Text Generation Kulikov, Ilia Abstract | PDF

  Title: Characterizing and Resolving Degeneracies in Neural Autoregressive Text Generation

  Candidate: Kulikov, Ilia

  Advisor(s): Kyunghyun Cho, Jason Weston

  Abstract:

  Autoregressive neural networks have shown great success as part of the sequence to sequence framework solving a diverse set of sequence generation tasks. These tasks include machine translation, dialogue modeling, question answering, text summarization, and sequence completion. In spite of the visible success, many challenges remain to be solved and are reported across these tasks. These challenges are usually discussed as visible deviations in the predicted sequence compared to the given reference. It is, however, not always possible to do the comparison, because interactive tasks, such as dialogue modeling, do not come together with reference sequences in the middle of the conversation at the test time. We refer to such deviations as \textit{degeneracies} which result in degenerate sequences. In this thesis, we work on reducing widely reported degeneracies within specific tasks or in text generation in general. To do so, we often first need to formulate the degeneracy in a measurable way and hypothesize what is the major cause behind it.
  We investigate the issue of oversmoothing, where the model assigns high probability to overly short sequences. We address this degeneracy from the learning aspect by proposing a novel regularization which minimizes the newly proposed oversmoothing rate directly. We show the effectiveness of the proposed method in the context of neural machine translation. Still concentrating on the learning aspect, we next address the problem of repetition in the context of sequence completion, where the generated sequences have unreasonably many repetitive substrings compared to the ones we see in the data. We propose a novel unlikelihood training procedure which allows to penalize undesired continuations, such as repetitive substrings. Unlikelihood training significantly reduces the number of repetitions and improves the naturalness of the generated continuations. One issue with the repetition degeneracy is that it can also lead to non-termination. We study if the original model is able to terminate the repetitive loop itself even if we do not enforce the maximum generated length during decoding. We connect this problem of non-termination with the consistency of the distribution induced by the chosen decoding algorithm. After proving that an incomplete decoding algorithm, such as beam search, may induce the inconsistent distribution when paired with a consistent model, we propose an alternative parametrization which guarantees the decoding-induced distribution to be consistent. After that, we switch to a more complicated scenario of conversation modeling, where the model has to generate a response in a multi-turn setting. We investigate the issue of unengaging or dull responses by highlighting the importance of the decoding algorithm. We observe a low diversity of beam search candidates compared to iterative beam search which explores a wider search subspace via efficient pruning. We find that the selection criterion is as important as the decoding strategy. Along the way, we stress the importance of careful human evaluation in the presence of annotator bias and calibrate the observed scores using Bayesian inference. While we address different kinds of degeneracy, the list we tackle is not exhaustive. For instance, neural machine translation is known to produce hallucinated translations or copy large parts of the input sentence. Furthermore, degeneracies exist past autoregressive modeling in both non-autoregressive and semi-autoregressive settings. We believe our contributions will be helpful for future research solving new problems.

   

• Ph.D. Thesis 2022 Finding and Fixing Undesirable Behaviors in Pretrained Language Models Perez, Ethan Abstract | PDF

  Title: Finding and Fixing Undesirable Behaviors in Pretrained Language Models

  Candidate: Perez, Ethan

  Advisor(s): Kyunghyun Cho

  Abstract:

  Natural Language Processing (NLP) promises to deliver tools for a variety of impactful applications, ranging from automatic summarization to question-answering systems and conversational assistants. Recently, NLP has been revolutionized by the advent of Pretrained Language Models (PLMs). We train PLMs using "self-supervised" learning objectives -- prediction tasks that operate on unlabeled text alone, such as next word prediction or missing word prediction. As a result, PLMs are able to learn from large quantities of internet text, to obtain strong performance on many NLP tasks.

  Despite the success of self-supervised objectives, they face a fundamental limitation: they train PLMs to behave in ways that are misaligned with human preferences. PLMs learn to repeat internet misinformation, offensive jokes, and personal contact information, and it is hard to control or guide the text that PLMs generate. Next, we show that PLM-based classifiers are effective at predicting which texts humans prefer. As a result, it is possible to use such classifiers as a learning signal to automatically correct the PLM. We showcase this approach to train a high-quality retrieval system, obtaining strong performance across a variety of tasks using Retrieval-Augmented Generation (RAG). Even after such training schemes, some undesirable behaviors may remain undetected during training. We thus go a step further and generate inputs that elicit undesirable behaviors from the PLM using other PLMs, to preemptively find and fix such behaviors. Overall, we find that some of the most powerful tools for aligning PLMs with human preferences are PLMs themselves.

• Ph.D. Thesis 2022 Identifying, Addressing, and Understanding Challenging Cases in Machine Learning Resnick, Cinjon Abstract | PDF

  Title: Identifying, Addressing, and Understanding Challenging Cases in Machine Learning

  Candidate: Resnick, Cinjon

  Advisor(s): Kyunghyun Cho/Joan Bruna

  Abstract:

  Machine learning has advanced tremendously this past decade. Object
  detection systems routinely perform beyond human-level accuracy with
  no loss in speed, game-playing agents play at superhuman level in real
  time, and generative models write language useful enough for
  downstream products. And yet, autonomous vehicles (AV) crash due to
  surprising mistakes, the best gaming agents lose to simple strategies,
  and our language models produce nonsensical utterances at a
  surprisingly high rate. I could have chosen examples from any field
  because these failures are not endemic to just vision, games, or
  language. There are always challenging cases remaining after training
  our system, and these cases are where the systems fail.

  This thesis focuses on the challenging cases in a machine learning
  system in order to improve its overall capabilities. In the first
  part, we study methods for identifying the challenging cases, an
  important precursor for improving the system. In the second part, we
  then study methods for addressing the challenging cases, arguably the
  most important part of this thesis for real-world applicability. And
  in the third part, we study methods for understanding the root cause
  of challenging cases, an important step in attaining guarantees about
  our system's capabilities. As machine learning is practiced in many
  different settings, our study does too. We explore these questions in
  the context of computer vision, language learning, and task learning.
  The connecting thread among them is the drive towards creating a
  communicative and visually aware robot that can capably complete
  household tasks. In that context, we present in parallel the Machine
  Learning Application Framework that highlights where our contributions
  improve downstream applications.

  All together, this work studies how to identify, address, and
  understand the most challenging cases over a diverse array of machine
  learning systems. This research is imperative towards deploying many
  systems that we care about, including most autonomous vehicles and
  health assistants. Consequently, it represents an important step
  towards society's technological goals.

• Ph.D. Thesis 2022 Constrained Surface Parameterization Methods with Guarantees Shen, Hanxiao Abstract | PDF

  Title: Constrained Surface Parameterization Methods with Guarantees

  Candidate: Shen, Hanxiao

  Advisor(s): Denis Zorin/Daniele Panozzo

  Abstract:

  Surface parameterization for piecewise-linear surfaces is a
  fundamental problem in computer graphics and geometry processing. The
  generation of surface parameterization is a key step in numerous
  applications like texture mapping, remeshing, quadrangulation,
  inter-surface mapping, and shape-analysis.

  Due to its popularity, the robustness of mapping
  generation methods plays a major role in its applicability. In
  addition, depending on the specific requirements of the application at
  hand, various formulations of constraints are used to control or guide
  the parameterization.
  Typical examples of the constraints are point constraints, curvature
  constraints, and topological constraints. In many practical cases, to
  ensure that the input assumptions of downstream algorithms are
  satisfied, the constraints, such as need to be imposed exactly (as
  opposed, e.g., to approximation via penalties).

  In this work, we investigate different
  constraint formulations suitable for various applications and present
  algorithms with guarantees to generate parameterization fully
  satisfying these constraints. In the first part of this thesis, we
  develop an algorithm that solves the
  classical problem of mapping a disk domain with boundary constraints;
  in the special case of domains with convex boundary, it improves, in
  terms of robustness, on the classical Tutte's algorithm. Utilizing it
  as a building block, we design a parameterization method that supports
  arbitrary positional constraints. In the second part, building on
  recent developments in the theory of discrete uniformization, we
  develop a highly robust algorithm for discrete conformal maps that
  satisfy prescribed curvature constraints. In the third part, we
  provide a constructive proof for the existence of globally seamless
  parameterization that matches admissible user-prescribed cone position
  and curvature constraints. Lastly, we generalize this to constraints
  on holonomy angles on a homology basis of loops, which fully capture
  the topology of seamless parameterizations. This method yields
  parameterizations that are very close to field-aligned
  parametrizations obtained using commonly used methods but, in contrast
  to these methods, guarantees the existence of solution satisfying all
  constraints.

• Ph.D. Thesis 2022 On deep learning tools for scientific discovery in healthcare Sudarshan, Mukund Abstract | PDF

  Title: On deep learning tools for scientific discovery in healthcare

  Candidate: Sudarshan, Mukund

  Advisor(s): Rajesh Ranganath/Oded Regev

  Abstract:

  Scientists validate hypotheses by building mathematical models of the
  real world. They make inferences by checking if their models are
  supported by data. Often, the models are hand-crafted and do not
  accurately reflect real processes. This often leads to low power to
  make scientific discoveries or even false discoveries.

  
  Machine learning can solve these issues in several ways. By allowing
  data to inform the construction of models, scientists can use machine
  learning to create more powerful statistical hypothesis testing
  procedures, or build more realistic models of underlying processes.
  
  This thesis details techniques to address both of these approaches.
  First we address the creation of machine learning-based statistical
  discovery procedures for scientific discovery. Specifically, we
  discuss how machine learning can be used to construct conditional
  independence tests, which are used to identify causal links in data.
  
  We detail how such methods can be used to control the false discovery
  rate when testing multiple hypotheses. We then apply these techniques
  to two important domains. We solve a timely problem in medical
  informatics: identifying a small set of variables that are highly
  informative of whether an ICU patient with Covid will experience an
  adverse event. At the height of Covid in 2020, NYU doctors used a
  deployed version of this tool to quickly identify patients to
  discharge and free up beds in the ICU. We also apply our methods to a
  problem in cancer genomics, where the goal is to identify a set of
  gene mutations that are most predictive of tumor metastasis. In the
  near future, we expect tools like ours to lead to targeted gene
  therapies that tailor treatments to the mutations present in an
  individual's tumor.

  
  Next we detail the construction of an interpretable machine learning
  model that helps understand an important step in the creation of
  proteins. Specifically, we build a model to understand RNA splicing,
  which involves removing non-coding regions from precursor messenger
  mRNA (pre-mRNA) and joining coding regions together. Our model
  accurately models splicing outcomes across a large dataset of
  sequences, but more importantly leads to several biologically
  validated insights. We use the interpretable nature of our model to
  infer that most splicing decisions are a function of a small set of
  short sequence features. We also learn that certain pre-mRNA secondary
  structures strongly inhibit the inclusion of a coding region in the
  final mRNA transcript. Finally, we validate these model-driven
  findings by carefully designing experiments for the wet lab.

   

• Ph.D. Thesis 2022 Efficient Verification of Untrusted Services Tzialla, Ioanna Abstract | PDF

  Title: Efficient Verification of Untrusted Services

  Candidate: Tzialla, Ioanna

  Advisor(s): Michael Walfish

  Abstract:

  Using a third-party service today requires trusting that it is executing as promised. Meanwhile, the correct execution of services is regularly impeded by failures, bugs, misconfigurations, operational mistakes, and insider attacks. Is it possible to verify, instead of trust, that a third-party service executes correctly?
  
  We study this question for two services that execute on remote servers: transparency dictionaries, a foundational infrastructure for end-to-end encryption and other applications, and event-driven web applications. For each of these two services, we leverage their workloads to introduce a practical system that allows a verifier to get a strong security guarantee that the service executes correctly.
  
  In the case of a transparency dictionary, this guarantee is in the form of a cryptographic proof provided by the service. Producing cryptographic proofs typically requires high resource costs. We show that tailoring the cryptographic tools used by the transparency dictionary for its use case mitigates these costs and results in a system, Verdict, that scales to dictionaries with millions of entries while imposing modest overheads on the service and its clients.
  
  In the case of outsourced event-driven web applications, the verifier gets the required guarantee by replaying the requests on a trusted machine using Karousos, a novel record-replay system in which the service has the role of the untrusted recorder. Karousos takes advantage of the particular characteristics of event-driven web applications to enable the replayer (the verifier) to use less computational resources than the recorder (the service), while imposing tolerable overheads on the
  recorder and keeping communication small.

• Ph.D. Thesis 2022 NLP Evaluation in the Time of Large Language Models Wang, Alex Abstract | PDF

  Title: NLP Evaluation in the Time of Large Language Models

  Candidate: Wang, Alex

  Advisor(s): Kyunghyun Cho

  Abstract:

  The field of natural language processing (NLP) has been
  >dramatically impacted by the creation and proliferation of large
   language models that are pretrained on Internet-scale text data. These
   models have led to significant improvements on a myriad of NLP tasks.
   However, as the capabilities of these models drive up performance on
   existing task benchmarks, there is a critical need for evaluation
   metrics that are up-to-date with current models. In this dissertation,
   we develop NLP evaluation methodologies that benchmark and leverage
   pretrained language models. We first present two multi-task benchmarks
   for evaluating the generalization ability of NLP models and discuss
   the role of these benchmarks in the development of large language
   models. Next, we demonstrate that we can leverage the capabilities of
   pretrained language models to develop new automatic evaluation metrics
   that better measure the semantics of model-generated text.
   Specifically, we make use of the question answering abilities of
   pretrained models to evaluate the faithfulness of automatically
   generated summaries. Finally, we explore methods for crowdsourcing
   high-quality and challenging text generation data to address issues of
   data quality that have been surfaced by the ability of language models
   to replicate noise in benchmark datasets. Overall, we show that the
   rise of pretrained language models presents both challenges and
   opportunities in how we evaluate NLP systems, and that incorporating
   these very models into our evaluation methodologies offers a promising
   path forward.

• Ph.D. Thesis 2022 Improving Sample Efficiency in Off-policy and Offline Deep Reinforcement Learning Wu, Yanqiu (Autumn) Abstract | PDF

  Title: Improving Sample Efficiency in Off-policy and Offline Deep Reinforcement Learning

  Candidate: Wu, Yanqiu (Autumn)

  Advisor(s): Keith Ross

  Abstract:

  Reinforcement Learning (RL) is an area of Machine Learning, where agents are trained through trial and error to make a sequence of decisions in some given environment to achieve a goal. Traditional reinforcement learning methodology suffers from the curse of dimensionality. Fortunately, with the help of deep learning, Deep Reinforcement Learning (DRL) can overcome the issue and can often find high performing policies for applications with large state and action spaces. Over the past few years, DRL has achieved major breakthroughs in complex tasks, such as outperforming human players in video games [Mnih et al. 2013; Vinyals et al. 2019], defeating the human world champion in Go [Silver et al. 2016, 2018] and autonomous robotics control [Lillicrap et al. 2019; Haarnoja et al. 2018a].
  
  Despite the recent breakthroughs, sample efficiency remains an important issue in deep reinforcement learning. In some complex tasks, where data collection is very expensive and agents require relatively few interactions with the environment for training, sample efficiency is of central concern for making DRL practical for applications. This thesis addresses the sample efficiency problem in the context of off-policy and offline Deep Reinforcement Learning. We develop training algorithms which not only lead to high asymptotic performing policies, but are also highly sample efficient in both on-line and offline settings. We demonstrate the performance of our methods in simulated robotic locomotion environments.
  
  In the first part of this thesis, we develop a streamlined off-policy algorithm that utilizes an output normalization scheme and non-uniform sampling. We identify the squashing exploration problem and show how maximum entropy DRL [Haarnoja et al. 2018a,b] helps to resolve it. Based on our observation, we develop an alternative output normalization scheme to maximum entropy algorithms. We show that this normalization scheme can then be combined with non-uniform sampling, resulting in high performing policies. Next, we develop a simple off-policy algorithm that takes advantage of a high update-to-data (UTD) ratio and Q-ensembles which demonstrates superior sample efficiency in early-stage training and also achieve high asymptotic performance in late-stage training. We employ Q-ensembles and keep several lowest values for updating to address the overestimation bias. Finally, we consider offline deep reinforcement learning. We introduce the novel notion of “upper envelope of the data” and then develop an Imitation-Learning based algorithm based on the notion. Our algorithm is computationally much faster and achieves state-of-the art performance.
   

• Ph.D. Thesis 2022 On-Policy Deep Reinforcement Learning — The Discounted and Average Reward Criteria Zhang, Yiming Abstract | PDF

  Title: On-Policy Deep Reinforcement Learning — The Discounted and Average Reward Criteria

  Candidate: Zhang, Yiming

  Advisor(s): Keith Ross

  Abstract:

  Reinforcement Learning (RL) is the study of sequential decision making where an agent attempts to maximize its overall cumulative reward in some given environment. Combined with deep learning, reinforcement learning has made remarkable strides in the past decade in complex tasks such as playing video games (Mnih et al. 2013, Vinyals et al. 2019), playing Go (Silver et al. 2016, 2018), robotics (Lillicrap et al. 2016, Haarnoja et al. 2018), and chip design (Mirhoseini et al. 2021). However despite these successes, modern RL algorithms often suffer from poor sample efficiency and lack of safety guarantees. In this thesis we tackle these issues in the context of on-policy Deep Reinforcement Learning (DRL), both theoretically and algorithmically. This work addresses both the discounted and average reward criteria. In the first part of this thesis, we develop theory for average reward on-policy reinforcement learning by extending recent results for local policy search. We show that previous work based on the discounted return (Schulman et al. 2015, Achiam et al. 2017) results in a non-meaningful bound in the average-reward setting. By addressing the average-reward criterion directly, we derive a novel bound which depends on the average divergence between the two policies and Kemeny's constant. Based on this bound, we develop an iterative procedure which produces a sequence of monotonically improved policies for the average reward criterion. We show that this iterative procedure can then be combined with classic deep reinforcement learning methods, resulting in practical DRL algorithms that target the long-run average reward criterion. Next, we develop a unifying framework for the on-policy sample efficiency problem. This methodology uses a two-step approach which first learns an optimal policy in the non-parameterized policy space before projecting said policy back into the parameter space. Our approach is general in that it applies to both discrete and continuous action spaces, and can handle a wide variety of proximity constraints. Finally we address the problem of reinforcement learning with safety constraints. We provide theoretical support that trust region-based methods can be applied to problems with both discounted and non-discounted cost constraints. We then propose a novel first-order algorithm for policy optimization for maximizing an agent's cumulative reward while at the same time satisfying a set of cost constraints. Our algorithm is extremely simple to implement and has an approximate upper bound for worst-case constraint violation throughout training.

• Ph.D. Thesis 2021 Advances in computer bridge: techniques for a partial-information, communication-based game. Bethe, Paul Abstract | PDF

  Title: Advances in computer bridge: techniques for a partial-information, communication-based game.

  Candidate: Bethe, Paul

  Advisor(s): Ernest Davis

  Abstract:

  Bridge is an imperfect information game with elements of competition
  against opponents as well as cooperation with a partner. Despite the
  application of many tenets of artificial intelligence, humans have yet
  to be consistently bested by the computer. This thesis explores AI
  shortcomings in both the play and bidding phases of the game. In the
  play, we explore weaknesses in the cutting edge Monte Carlo techniques
  and explore both inference and learning based solutions. In the bidding,
  we go beyond existing rule based systems and investigate deep
  reinforcement learning as a method to learn how to bid.

• Ph.D. Thesis 2021 Learning Causality in Molecular Biology Cirrone, Jacopo Abstract | PDF

  Title: Learning Causality in Molecular Biology

  Candidate: Cirrone, Jacopo

  Advisor(s): Dennis Shasha

  Abstract:

  The Systems Biology community has invested a great deal of effort in
  modeling gene regulatory networks that should be able to (i) accurately
  predict future states and (ii) identify regulatory hubs that can be
  manipulated to achieve desired phenotypes. Most computational tools for
  the problem embody linear models (e.g. 5 * TF1 + 2*TF2 - 0.4*TF3....).
  However, it is well known that biological interactions are highly
  synergistic and non-linear. Further, those tools mostly try to directly
  predict networks even when the discovered edges (which usually come from
  some assay such as Chip-seq) may have little physiological significance
  (e.g., may not influence gene expression).

  This thesis considers an alternative approach to inferring gene
  causality. Specifically, we consider the problem of predicting the
  expression of genes at a future time point in a genomic time series. In
  this, we follow the philosophy that accurate prediction often
  corresponds to a good understanding of causality.
  The prediction may rest on several sources of data: the time point
  immediately preceding t, the entire target time series preceding t, and
  ancillary data. In biology, for example, the ancillary data may consist
  of a network based on binding data, data from different time series,
  steady state data, a community-blessed gold standard network, or some
  combination of those. We introduce OutPredict, which is a machine
  learning method for time series that incorporates ancillary steady state
  and network data to achieve a low error in gene expression prediction.
  We show that OutPredict outperforms several of the best state-of-the-art
  methods for prediction. The predictive models OutPredict in turn
  generate a causal network.

  Thus, this thesis presents an approach to the inference of causality
  based on predictions of out-of-sample time-points based on both steady
  state and time series data. Because the model for each gene identifies
  those transcription factors that have the most importance in prediction,
  those important transcription factors are the most likely causal
  elements for that gene. We validate those predictions for a set of
  well-documented transcription factors in Arabidopsis.
  Because our methods apply to any situation in which there is time series
  data, ancillary data, and the need for non-linear causal models, we
  believe that this work will have a broad appeal to the scientific
  community, specifically those studying causality networks in any
  biological system.

   

• Ph.D. Thesis 2021 Responsibility Analysis by Abstract Interpretation Deng, Chaoqiang Abstract | PDF

  Title: Responsibility Analysis by Abstract Interpretation

  Candidate: Deng, Chaoqiang

  Advisor(s): Patrick Cousot

  Abstract:

  Given a behavior of interest, automatically determining the corresponding responsible entity (or say, the root cause) is a task of critical importance in various scientific fields, especially in the program static analysis. Classical static analysis techniques (e.g. dependency analysis, taint analysis, slicing, etc.) assist programmers in narrowing down the scope of responsibility, but none of them can explicitly identify the responsible entity. Meanwhile, the causality analysis is generally not pertinent for analyzing programs, and the structural equations model (SEM) of actual causality misses some information inherent in programs (e.g. temporal information, and whether an entity is free to make choices or not), making the corresponding program analysis imprecise.

  In this dissertation, inspired by a classic forest fire example used in defining causality, a novel definition of responsibility based on the abstraction of trace semantics is proposed, which is expressive and generic to cope with both program analyses and tasks in other scientific fields. Briefly speaking, an action aR is responsible for behavior B in a certain trace, if and only if aR is free to make choices, and such a choice is the first one that ensures the occurrence of B in that trace. Such a definition makes use of the information regarding the temporal ordering of actions, as well as whether an action has free choices or not. In addition, our definition of responsibility takes into account the cognizance of observer, which, to the best of our knowledge, is a new innovative idea in program analysis. Compared to current dependency and causality analysis methods, the responsibility analysis is demonstrated to be more precise in many examples.

  Furthermore, this dissertation proposes a sound framework of abstract responsibility analysis, which allows a balance between cost and precision to solve the undecidable problem of responsibility. Essentially, the abstract analysis builds a trace partitioning automaton by an iteration of over-approximating forward reachability analysis with trace partitioning and under-approximating/over-approximating backward impossible failure accessibility analysis, and determines the bounds of potentially responsible entities along paths in the automaton. Unlike the concrete responsibility analysis identifies exactly a single action as the responsible entity along every concrete trace, the abstract analysis may lose some precision and find multiple actions potentially responsible along each automaton path. However, the soundness is preserved, and every responsible entity in the concrete is guaranteed to be also found responsible in the abstract.

   

• Ph.D. Thesis 2021 Enhancing Collaboration and Productivity for Virtual and Augmented Reality He, Zhenyi Abstract | PDF

  Title: Enhancing Collaboration and Productivity for Virtual and Augmented Reality

  Candidate: He, Zhenyi

  Advisor(s): Ken Perlin

  Abstract:

  Immersive environments such as Virtual Reality (VR) and Augmented Reality (AR) are now receiving more and more attention. Although VR and AR have largely been used for individual entertainment experiences, they also possess huge potential as a platform for the support of collaboration and productivity. My thesis work is concerned with enabling VR/AR to be flexibly adapted for collaborative and productive uses. I approach this scope from several facets: a new haptic user interface based on actuated robots to bridge virtual and physical world, a reconfigurable framework for both co-located and geographically dispersed multi-user communication, and a text entry system in which users type by tapping their fingers, without needing to look at their hands or be aware of their hand positions. Further, I extend these ideas to a daily video conferencing experience that requires minimal hardware.

• Ph.D. Thesis 2021 Larger-Context Neural Machine Translation Jean, Sébastien Abstract | PDF

  Title: Larger-Context Neural Machine Translation

  Candidate: Jean, Sébastien

  Advisor(s): Kyunghyun Cho

  Abstract:

  Translation helps connect people by bridging language barriers. It can make travel more enjoyable, allow our minds to explore imaginary worlds, let us talk to others, and so on. Given the need for translation, but the limited availability of human translators, machine translation has flourished. Most systems translate sentences one by one, ignoring its context, which isn't always sufficient as the missing information can lead to incorrect or inconsistent translations. We believe that neural machine translation (NMT) is particularly well-suited to incorporate the surrounding context. Indeed, NMT systems can attend to arbitrarily distant words, while the use of continuous representations improves generalization on unseen examples.

  As such, in this thesis, we extend neural machine translation to leverage information from the surrounding context. To do so, we first highlight the potential of the then-nascent NMT paradigm. We subsequently introduce architectural changes to integrate information from the surrounding document, initially starting from the preceding sentence. We further encourage models to use context from either a learning or data augmentation perspective. We also consider the efficient use of document-level neural language models for this task. While some challenges still remain, our work has helped establish larger-context translation on a solid footing, and we are optimistic about future progress.

• Ph.D. Thesis 2021 Improving Sample Efficiency of Imitation and Reinforcement Learning Kostrikov, Ilya Abstract | PDF

  Title: Improving Sample Efficiency of Imitation and Reinforcement Learning

  Candidate: Kostrikov, Ilya

  Advisor(s): Rob Fergus

  Abstract:

  Reinforcement Learning (RL) is an area of machine learning focused on learning to make a sequence of actions in an environment that maximizes cumulative rewards. Combined with Deep Learning, Reinforcement Learning has made significant progress over the last decade across various domains. Notable successes include achieving superhuman performance on Atari games, Go, StarCraft II, Dota 2, and various continuous control tasks.

  However, RL's success stories are often limited to games and simulations where it is possible to generate a large amount of training data. This thesis describes several methods focused on improving sample efficiency to enable a wider variety of RL applications. For the first half of the thesis, we focus on Imitation Learning, where ground truth rewards are usually unknown, and expert demonstrations define optimality. First, we introduce a method for robust and sample efficient imitation learning. We adapt an imitation learning approach where an agent tries to mimic a domain expert using a GAN-like framework called GAIL. We identify two primary sources of sample inefficiency associated with this approach: on-policy RL and GAN discriminator training. We show that sample inefficiency can be mitigated by performing off-policy RL training combined with off-policy training of the discriminator. We also identify and resolve some task-specific biases associated with the family of adversarial imitation learning algorithms based on GAIL. Then, we derive a principled off-policy formulation of robust imitation learning that is entirely offline and allows us to learn a policy that imitates the expert relying only on the previously collected data. This work concludes the part of the thesis focused on imitation learning, and for the rest of the thesis, we focus on online and offline RL where we have access to environment rewards. We observe that off-policy RL from pixels suffers from overfitting and propose a simple solution inspired by image augmentation techniques from Computer Vision. Finally, we introduce a method for offline RL that utilizes a pre-trained behavioral policy to improve the robustness of behavior regularization widely used in the context of offline RL. In contrast to prior work on Offline RL, this method utilizes the behavior policy to regularize the critic instead of constraining the training policy.

• Ph.D. Thesis 2021 Latent Variable Models and Iterative Refinement for Non-Autoregressive Neural Machine Translation Lee, Jason Abstract | PDF

  Title: Latent Variable Models and Iterative Refinement for Non-Autoregressive Neural Machine Translation

  Candidate: Lee, Jason

  Advisor(s): Kyunghyun Cho

  Abstract:

  Deep neural networks have fundamentally transformed the field of machine translation, and replaced statistical phrase-based approaches to serve translations to millions of users in production systems every day. Despite impressive progress in translation accuracy, improving decoding speed remains a key challenge as most systems are \emph{autoregressive} and generate a sentence word-by-word. As neural machine translation (NMT) models are becoming increasingly deep and complex, there is a growing need for more efficient translation systems with sub-linear or constant inference latency, with respect to the sentence length. The main challenge in non-autoregressive machine translation is capturing the dependencies between tokens in a target sentence without autogression. Motivated by a rich history of probabilistic graphical models in sequence generation, this thesis proposes to use latent variables to model intra-sentence dependencies, such that the output distribution can be factorized given the latent variables. We also present several inference algorithms for non-autoregressive machine translation based on iterative refinement, which revises a sentence over multiple iterations. Our non-autoregressive models based on latent variables and iterative refinement can deliver significant decoding speedup with comparable translation accuracy relative to a strong autoregressive baseline. Finally, we investigate the correlation between training (log-likelihood) and test objective (BLEU) of several model families. We observe the two metrics are not correlated when comparing models from different families (e.g. between autoregressive and latent variable models).

• Ph.D. Thesis 2021 Neural Structured Prediction using Iterative Refinement with Applications to Text and Molecule Generation Mansimov, Elman Abstract | PDF

  Title: Neural Structured Prediction using Iterative Refinement with Applications to Text and Molecule Generation

  Candidate: Mansimov, Elman

  Advisor(s): Kyunghyun Cho

  Abstract:

  Humans excel at generating structured data in the form of images, text, speech, molecules, computer code, and others. Researchers have spent several decades proposing various solutions for the effective generation of these structured objects in a data-driven way, known as structured prediction. With the revival of deep neural networks, autoregressive models that process structured objects in fixed left-to-right monotonic ordering became a de-facto solution for this problem. Notable successes of autoregressive models include neural machine translation [Sutskever et al., 2014, Bahdanau et al., 2014, Vaswani et al., 2017], open-ended text generation [Radford et al., 2019, Brown et al., 2020], text-to-speech synthesis [van den Oord et al., 2016], among many.

  Despite the considerable success of autoregressive models on many applications, a natural question arises whether alternative approaches are possible for structured prediction. This thesis describes a novel method for structured prediction based on the principle of iterative refinement with a particular focus on applications to text and molecule generation. We first introduce the iterative refinement framework for text generation. Starting from the blank sentence, the iterative refinement approach gradually refines text over multiple steps. Using this approach, we show that we can flexibly generate the text in various ways, such as generate all or some words in parallel and generate text according to the ordering learned from the data. We show that iterative refinement achieves competitive performance compared to autoregressive models while delivering a speedup in decoding. We conclude this thesis by showing how we can adapt the iterative refinement framework originally introduced for text generation for molecule generation. In particular, we demonstrate two iterative refinement approaches for molecular graph generation and molecular geometry prediction. We anticipate that models based on the iterative refinement will be broadly applicable to other domains of interest.

• Ph.D. Thesis 2021 Scalable Particulate Flow Simulations with Boundary Integral Equations Morse, Matthew Abstract | PDF

  Title: Scalable Particulate Flow Simulations with Boundary Integral Equations

  Candidate: Morse, Matthew

  Advisor(s): Denis Zorin

  Abstract:

  Numerical simulation of complex particulate flows, and of red blood cell flows through capillaries in particular, is an important investigational tool in the biological sciences. The ability to rapidly evaluate the impact of vessel and cell geometries, plasma viscosity, and particulate densities on macroscopic physiology is crucial to pursuing further biological understanding. Experimental techniques are costly and time-consuming, while analytical approaches are often of limited practical use in realistic scenarios, ultimately underscoring the importance of a comptuational approach.

  In this work, we construct such a simulation, capable of simulating microliters of blood flowing through realistic vasculature, along with more general particulate suspensions. Due to the micrometer length scales of typical capillaries, we can model the blood plasma as a Stokesian fluid and red blood cells as inextensible, deformable membranes. By reformulating the viscous flow as a set of boundary integral equations, we are able to produce a method that has optimal complexity with high-order accuracy that is capable of handling dense particulate suspensions in complex geometries.

  This approach relies on a novel, robust solver for elliptic partial differential equations, applied to Stokes flow. A core component of the solver is a novel fast algorithm to compute the value of the solution near and on the domain boundary, which we have named \qbkix. We provide a set of algorithms to guarantee the accuracy of \qbkix on piecewise smooth surfaces, discuss the error behavior and complexity of \qbkix, and evaluate its performance.

  Leveraging this solver in a confined blood flow simulation involves advecting deformable particulates along the flow trajectory. Large timesteps are required for an efficient simulation, but can cause collisions among cells and with the vessel wall if performed naively. We present collision detection and resolution algorithms for the red blood cells and the blood vessel. We parallelize \qbkix and the collision algorithms and scale the final simulation to nearly 35,000 cores.

• Ph.D. Thesis 2021 Towards More General and Adaptive Deep Reinforcement Learning Agents Raileanu, Roberta Abstract | PDF

  Title: Towards More General and Adaptive Deep Reinforcement Learning Agents

  Candidate: Raileanu, Roberta

  Advisor(s): Rob Fergus

  Abstract:

  Building agents with general skills that can be applied in a wide
   range of settings has been a long-standing problem in machine
   learning. The most popular framework for training agents to make
   sequential decisions in order to maximize reward in a given
   environment is Reinforcement Learning (RL). Over the last decade, deep
   reinforcement learning, where RL agents are parameterized by neural
   networks, has achieved impressive results on a number of tasks, from
   games such as Atari, Go, StarCraft, or Dota, to continuous control
   tasks with applications in robotics._

   However, current RL agents are prone to overfitting and struggle to
   generalize when even minor perturbations are applied to the training
   environment. This hinders progress on real-world applications such as
   autonomous vehicles or home robots, where agents need to deal with a
   large variety of scenarios. In this thesis, we introduce several
   methods for improving the versatility of deep reinforcement learning
   agents. We start by studying the problem of zero-shot generalization
   to new instances of a task after training on a limited number of
   environments. We first propose an approach for regularizing the policy
   and value function of a RL agent and automatically finding an
   effective type of data augmentation for a given task. We also identify
   that there is an asymmetry between the information needed to represent
   the optimal policy and the true value function, which leads to
   overfitting when using standard deep RL algorithms. As a step towards
   solving this problem, we propose a method which decouples the
   optimization of the policy and value, and constrains the
   representation to be invariant to the task instance. Next, we focus on
   the problem of learning general exploration strategies for
   procedurally generated environments with sparse rewards. We formulate
   a new type of intrinsic reward which encourages agents to impact their
   environments and show that it outperforms other popular exploration
   methods. Then, we discuss a novel approach for fast adaptation to new
   dynamics. We show that our method, which leverages self-supervised
   techniques to learn policy and environment embeddings, enables
   adaptation within a single episode on a number of continuous control
   tasks. Finally, we investigate how agents can learn more flexible
   strategies for interacting with different opponents and collaborators.

• Ph.D. Thesis 2021 Theory and Algorithms for Several Central Problems in Large-Scale Machine Learning Storcheus, Dmitry Abstract | PDF

  Title: Theory and Algorithms for Several Central Problems in Large-Scale Machine Learning

  Candidate: Storcheus, Dmitry

  Advisor(s): Mehryar Mohri

  Abstract:

  This Ph.D. dissertation presents fundamental analaysis of several central problems in large-scale machine learning. We derive novel, scalable algorithms supported by strong theoretical guarantees for the most practically important large-scale learning scenarios. These scenarios include extentions of the standard supervised learning to multiple base hypotheses spaces, multiple objective functions, multiple distributions, multiple classes and high-dimensional feature spaces.

  A standard supervised learning scenario consists of fitting a predictor from a fixed hypotheses space that minimizes certain empirical loss on a sample drawn i.i.d. from a particular distribution. The richness of modern machine learning applications requires the learning scenario to be large-scale by having the ability to learn from many training examples. While scalability in terms of many examples is widely studied, the current state of research in the field overlooks other scenarios and directions for scalability that may be even more important that many training examples. For instance, by allowing the learner to select predictors from multiple hypotheses spaces of varying complexity, or fit to multiple objective functions.

  While the problems mentioned above may seem to relate to separate aspects of large-scale learning, this thesis provides a unified theoretical analysis framework that brings these central problems together. This framework is based on the Rademacher complexity analysis as well as on the Empirical and Structural Risk Minimization principles.

• Ph.D. Thesis 2021 The Evolutionary Maps of Data Tamaskar, Abhinav Abstract | PDF

  Title: The Evolutionary Maps of Data

  Candidate: Tamaskar, Abhinav

  Advisor(s): Bud Mishra

  Abstract:

  We present a geometric view of analyzing temporal causal models from the perspective of topology and limit graphs. We will briefly cover an intuitive overview of the topological techniques used and the theory of limit graphs. We will then briefly describe the Suppes Bayes causal networks which are used as the temporal causal models. We briefly describe evolutionary models used in scientific literature and show an efficient model for performing simulations on generalized large scale evolutionary networks. We then present the techniques for analyzing large scale evolutionary populations, and showcase their generality through two real world examples, (1) with the linguistic data from Reddit over the course of 5 years and while showing the existence of echo chambers and giving a metric to analyze similarities of populations over time, and (2) through the TCGA and COSMIC dataset for cancer mutation of over 11,000 genes and by using an approximation metric on the space of causal models to find similar cancer types, to perform transfer learning to boost survival forecasting through blackbox learning models.

• Ph.D. Thesis 2021 Order and Learning in Sequential Neural Structured Prediction Welleck, Sean Abstract | PDF

  Title: Order and Learning in Sequential Neural Structured Prediction

  Candidate: Welleck, Sean

  Advisor(s): Kyunghyun Cho

  Abstract:

  Structured objects such as sets, trees, and sequences appear in a variety of scientific and industrial domains. Developing machine learning methods that generate these objects is of interest for both scientific understanding and practical applications. One approach, sequential neural structured prediction, decomposes generation into a sequence of predictions, with each prediction made by a deep neural network. Choosing an appropriate sequential representation of each structured object and selecting an effective learning objective are key to adopting this approach. The standard method for learning specifies a canonical ordering of elements in the sequential representation and maximizes the likelihood of the resulting sequences. We develop two streams of research that explore alternatives to this fixed-order, maximum likelihood approach for sequentially generating sets, trees, and sequences, with a focus on natural language processing applications.

  
  First, we focus on text generation and study degenerate properties of fixed-order maximum-likelihood learning, motivating new learning methods. We characterize the degeneracy using three properties observed in generated text: non-termination, logical incoherence, and repetition. To study non-termination, we develop theory that allows us to prove that conventional text generation methods can generate infinite-length sequences with high probability. To study logical incoherence, we create a dataset for investigating the degree to which a model logically contradicts its preceding statements. For reducing degeneration, we develop unlikelihood training, a learning method which penalizes task-specific textual properties. In the second part of the thesis, we remove the requirement of a fixed generation order with a learning framework called non-monotonic generation, which yields models that select input-dependent generation orders. We use non-monotonic generation to generate multisets, parse trees, and text. The investigations and techniques presented in this thesis lead to promising directions for future work.

• Ph.D. Thesis 2021 Techniques for Sample-Efficient Reinforcement Learning Whitney, William Abstract | PDF

  Title: Techniques for Sample-Efficient Reinforcement Learning

  Candidate: Whitney, William

  Advisor(s): Kyunghyun Cho

  Abstract:

  By leveraging advances in deep learning, reinforcement learning (RL) has recently made such advances that for any task which has a simulator, and thus enables the collection of nearly unlimited data, it might now be expected to yield superhuman performance. However, many practically relevant tasks take place in the physical world. Constructing physical simulators of sufficient fidelity and correspondence to transfer is a non-trivial challenge, so for the majority of physical tasks at least some amount of training on real data is required. Collecting data in the real world is sufficiently expensive that it makes up much of the cost of training a reinforcement learning agent.

  This thesis focuses on improving the sample efficiency of reinforcement learning in order to make them more practical to use on physical systems. It includes three approaches to this goal. The first part studies the data collection process, and in particular the opportunity for exploration to improve the sample efficiency of RL. The second part considers the use of representation learning to improve generalization, and thus sample efficiency, in reinforcement learning. The third part examines the offline RL setting, which consists of pure policy optimization using a fixed dataset and therefore does not require additional data collection.

  Taken together, this work studies techniques for improving the sample efficiency of reinforcement learning by collecting data which is more useful and diverse, then learning more from every sample.
  It represents an early step on the path to RL as an everyday tool for control of physical systems.

• Ph.D. Thesis 2021 Methods to Improve Knowledge Transfer Efficiency for Data-limited Problems in Genomics Yi, Ren Abstract | PDF

  Title: Methods to Improve Knowledge Transfer Efficiency for Data-limited Problems in Genomics

  Candidate: Yi, Ren

  Advisor(s): Richard Bonneau

  Abstract:

  The recent advancement in computational genomics has greatly benefited from the explosion of high-throughput genomic data and similar growth in biological databases. However, as more sequencing technologies become available and large genomic consortiums start to crowdsource data from larger cohorts of research groups, data heterogeneity has become an increasingly prominent issue. Data integration across multiple data sources and data modalities becomes particularly important for a greater number of biological systems. High-throughput omics data are typically highly skewed towards a small number of model organisms, factors, and conditions with which wet-lab experiments have higher success rates. It further introduces technical challenges when building machine learning models for problems with limited data. This thesis describes methods that improve knowledge transfer efficiency for learning data-limited problems through effective task-specific feature representation in the multitask learning setting. We demonstrate the performance of our methods in two genomic problems -- genetic variant calling and cell type-specific transcription factor binding predictions.

• Ph.D. Thesis 2020 Out of Distribution Generalization in Machine Learning Arjovsky, Martin Abstract | PDF

  Title: Out of Distribution Generalization in Machine Learning

  Candidate: Arjovsky, Martin

  Advisor(s): Leon Bottou

  Abstract:

  Machine learning has achieved tremendous success in a variety of
  domains in recent years. However, a lot of these success stories have
  been in places where the training and the testing distributions are
  extremely similar to each other. In everyday situations when models
  are tested in slightly different data than it was trained on, ML
  algorithms can fail spectacularly. This research attempts to formally
  define this problem, what sets of assumptions are reasonable to make
  in our data and what kind of guarantees we hope to obtain from them.
  Then, we focus on a certain class of out of distribution problems,
  their assumptions, and introduce simple algorithms that follow from
  these assumptions, and that are able to provide more reliable
  generalization. A central topic in the thesis is the strong link
  between discovering the causal structure of the data, finding features
  that are reliable (when using them to predict) regardless of their
  context, and out of distribution generalization.

   

• Ph.D. Thesis 2020 Behavior of the Limited-Memory BFGS Method on Nonsmooth Optimization Problems Asl, Azam Abstract | PDF

  Title: Behavior of the Limited-Memory BFGS Method on Nonsmooth Optimization Problems

  Candidate: Asl, Azam

  Advisor(s): Michael Overton

  Abstract:

  The limited memory BFGS (Broyden-Fletcher-Goldfarb-Shanno) method,
  abbreviated L-BFGS, is widely used for large-scale unconstrained optimization, but its behavior on nonsmooth problems has received little attention. In this thesis we give the first convergence analysis of the L-BFGS method applied to nonsmooth functions. We focus on the simplest version of the method, sometimes known as memoryless BFGS, which uses just one update. L-BFGS can be used with or without “scaling”; the use of scaling is normally recommended. We consider a simple class of convex piecewise linear nonsmooth functions f that are unbounded below. On this class of problems, we show that memoryless BFGS with scaling, using any ArmijoWolfe line search and initialized at any point where f is differentiable, generates iterates that converge to a non-optimal point, if a certain condition
  relating the Lipschitz constant of f to the line search Armijo parameter holds. We also present an analysis of the ordinary gradient method with the same line search applied to the same class of functions, giving conditions under which it also fails. However, scaled memoryless BFGS fails under a weaker condition relating the Lipschitz constant of the function to the line search Armijo parameter than that implying failure of the gradient method. Furthermore, in sharp contrast to the gradient method, if a specific standard Armijo-Wolfe bracketing line search is used, scaled memoryless BFGS fails if the Lipschitz constant is sufficiently large regardless of the Armijo
  parameter. Our experimental results demonstrate that our analysis is tight on this class of functions, and that similar results likely hold for L-BFGS with any fixed number of updates. In contrast, the “full” BFGS method is remarkably effective for minimizing nonsmooth functions, but it is not a practical approach when the number of variables is large.
  
  We also conduct extensive experiments applying L-BFGS, both unscaled, with various choices for the number of updates, on many other classes of convex nonsmooth functions, ranging from artificially devised, highly ill-conditioned nonsmooth problems to eigenvalue optimization problems that are equivalent to semidefinite programming problems arising from applications. We also apply L-BFGS to smoothed versions of these problems. We find that although L-BFGS is usually a reliable method for minimizing ill-conditioned smooth problems, when the condition number is so large that the function is effectively nonsmooth, L-BFGS consistently fails. This behavior is in sharp contrast to the behavior of full BFGS, which is consistently reliable for nonsmooth optimization problems. We arrive at the conclusion that, for large-scale nonsmooth optimization problems for which BFGS and other methods are not practical, it is far preferable to apply L-BFGS to a smoothed variant of a nonsmooth problem than to apply it directly to the nonsmooth problem.

• Ph.D. Thesis 2020 Enhanced Representations for Relations by Multi-task Learning Fu, Lisheng Abstract | PDF

  Title: Enhanced Representations for Relations by Multi-task Learning

  Candidate: Fu, Lisheng

  Advisor(s): Grishman, Ralph

  Abstract:

  A relation describes the relationship between a pair of entities. Relation Extraction is the process of extracting relations from free text and converting them to structured machine-readable knowledge. This process can facilitate building and extending knowledge bases, and therefore can benefit a variety of natural language processing applications such as Question Answering and Summarization.

  Typical relation extraction projects start by defining a relation schema:  a set of mutually-exclusive relation types.  Based on these definitions, all instances of these relations in a text corpus are labeled by hand, producing a dataset which can be used to train a statistical model. Labeling relations in text is difficult and  time-consuming. There only exist limited relation datasets developed in this way. New applications will give rise to new schemas, so the lack of high-quality labeled data is almost inevitable for Relation Extraction.

  Despite limited labeled samples in relation datasets, neural net models have been shown to be more effective than traditional methods in learning feature representations with pre-trained word embeddings. In the context of representation learning, this thesis presents multi-task learning frameworks to learn enhanced representations for relations. It shows how to learn better feature representations in both unsupervised and supervised ways. First, the dissertation shows how to learn domain invariant representations using unlabeled entity pairs. Then it shows how to learn a unified encoder by combining multiple annotated datasets. Finally, it shows how to learn the relatedness between relation types across different relation schemas. These techniques improve the relation models without requiring more annotation from the target dataset. The multi-task learning frameworks could be an efficient toolkit for relation extraction in general.

• Ph.D. Thesis 2020 Scaling Multi-user Virtual and Augmented Reality Herscher, Sebastian Abstract | PDF

  Title: Scaling Multi-user Virtual and Augmented Reality

  Candidate: Herscher, Sebastian

  Advisor(s): Perlin, Ken

  Abstract:

  The Virtual and Augmented Reality (XR) ecosystems have been gaining substantial momentum and traction within the gaming, entertainment, enterprise, and training markets in the past half-decade, but have been hampered by limitations in concurrent user count, throughput, and accessibility to mass audiences. Although a litany of XR devices have been made available for public purchase, most XR experiences have been developed for either a single user or a small set of users at a time. Few systems or experiments in co-located XR environments have expanded past a small set of users, leaving the paradigm of being part of a larger virtual audience relatively untested. This thesis presents a set of components, systems, and experiments that assist in the creation, deployment, and scaling of multi-user virtual and augmented reality experiences, and outlines the strengths of techniques found in traditional co-located media for the design space of scaled co-located XR.

• Ph.D. Thesis 2020 Auditing Outsourced Services Tan, Cheng Abstract | PDF

  Title: Auditing Outsourced Services

  Candidate: Tan, Cheng

  Advisor(s): Michael Walfish

  Abstract:

  Outsourcing to the cloud is based on assuming that remote servers behave as expected, even under failures, bugs, misconfigurations, operational mistakes, insider threats, and external attacks. Can we instead verify their behavior? There have been various attempts at such verification, but these attempts have had to choose: comprehensive guarantees or good performance? This dissertation studies how to get both.
  
  This dissertation focuses on two essential services: outsourced computation and outsourced databases. Verifying them correspondingly introduces two new abstract problems. We call the first problem the Efficient Server Audit Problem, which examines how to efficiently verify a concurrent and untrusted server. The second problem is verifying a core correctness contract of black-box databases while scaling to real-world online workloads.
  
  To address the two problems, this dissertation respectively introduces two systems: orochi and cobra. Both systems tolerate arbitrary failures in the service provider, and have good performance: in our experiments, orochi’s verifier achieves 5.6–10.9x speedup versus simply re-executing inputs, with less than 10% CPU overhead on the server side; cobra improves over baselines by 10x in verification cost, with modest overhead on clients (less than 5% throughput degradation and about 7% 90-percentile latency increases).

• Ph.D. Thesis 2020 Market Efficiency and Dynamics Tao, Yixin Abstract | PDF

  Title: Market Efficiency and Dynamics

  Candidate: Tao, Yixin

  Advisor(s): Richard Cole

  Abstract:

  General equilibrium theory, initiated by Walras over a century ago, explains the interaction between supply and demand in an economy. In this dissertation, we look at Fisher Markets, which are a particular case of the general equilibrium theory. We consider two issues in Fisher Markets: strategic behavior and dynamics.
  
  Strategic behavior is usually considered in a game, such as auction, in which case, participants in the game may choose not to report their real preferences in order to improve their payoff. In general equilibrium theory, buyers are usually considered to be non-strategic: given the prices, buyers will maximize their true utilities by properly distributing their money on different goods. In this case, the Market equilibrium should be efficient. However, the prices in the market equilibrium are influenced by the demands of the buyers. In principle, buyers can affect prices by changing their demands, which may improve buyers' final utilities. This may result in inefficient outcomes. In this thesis, we investigate this possibility in large Fisher markets. We show that the market will approach full efficiency as the market becomes larger and larger. We also show a similar result for the Walrasian mechanism in large settings.
  
  We also study two dynamics in Fisher Markets in this dissertation:

  1. Proportional response is a buyer-oriented dynamics. Each round, buyers update their spending in proportion to the utilities they received in the last round, where prices are the sum of the spendings. This dissertation establishes new convergence results for two generalizations of proportional response in Fisher markets with buyers having CES utility functions. The starting points are respectively a new convex and a new convex-concave formulation of such markets. The two generalizations of proportional response correspond to suitable mirror descent algorithms applied to these formulations. Among other results, we analyze a damped generalized proportional response and show a linear rate of convergence in a Fisher market with buyers whose utility functions cover the full spectrum of CES utilities aside the extremes of linear and Leontief utilities; when these utilities are included, we obtain an empirical O(1 / T) rate of convergence.
  2. Tatonnement is considered the most natural dynamics in Fisher Markets: the price of a good is raised if the demand exceeds the supply of the good, and decreased if it is too small. Implicitly, buyers' demands are assumed to be a best-response to the current prices. This dissertation addresses a lack of robustness in existing convergence results for discrete forms of tatonnement, including the fact that it need not converge when buyers have linear utility functions. This dissertation shows that for Fisher markets with buyers having CES utility functions, including linear utility functions, tatonnement will converge quickly to an approximate equilibrium (i.e., at a linear rate), modulo a suitable large market assumption. The quality of the approximation is a function of the parameters of the large market assumption.
• Ph.D. Thesis 2020 Flexible and Efficient Systems for Training Emerging Deep Neural Networks Wang, Minjie Abstract | PDF

  Title: Flexible and Efficient Systems for Training Emerging Deep Neural Networks

  Candidate: Wang, Minjie

  Advisor(s): Li, Jinyang

  Abstract:

  The success of deep neural networks (DNNs) is due to its strong capability to learn from data. To leverage more data requires larger models that may exceed the capacity of a single computing device. To leverage graph structured data demands models of sparse computation pattern. Unfortunately, current deep learning systems limit the exploration of such models, causing disturbing user experience. This thesis proposes a system design to guide the development of new deep learning systems. The goal of this design is to enable efficient training of these emerging DNNs with little user effort.

  We then realize the design in two systems, Tofu and DGL. Tofu partitions very large DNNs across multiple GPUs to reduce per-GPU memory footprint. To automatically partition each operator, we propose a description language for annotating the semantics of an operator. To optimally partition the whole training, Tofu proposes an algorithm that minimizes the total communication cost. We evaluate and assess the capability of Tofu to train very models demonstrating the substantial gains by applying the design. We then implement DGL, a new framework for training DNNs for graph structured data. DGL provides an intuitive and expressive interface that can cover a wide range of graph DNN models. We introduce batching and kernel fusion techniques that enable training GNNs on large graphs and achieve significant improvements in performance relative to existing systems.

• Ph.D. Thesis 2019 From 2.5G To 5G: Enhancing Access And Performance For Mobile Users Ahmad, Talal Abstract | PDF

  Title: From 2.5G To 5G: Enhancing Access And Performance For Mobile Users

  Candidate: Ahmad, Talal

  Advisor(s): Subramanian, Lakshminarayanan

  Abstract:

  This dissertation has two overarching themes: i) enhancing connectivity access for mobile users in rural contexts and ii) enhancing transport layer performance for mobile users.

  More than half of the world’s population faces barriers in accessing the Internet. A recent ITU study estimates that 2.6 billion people cannot afford connectivity and that 3.8 billion do not have access to it. To enhance access I have worked on two projects: Wi-Fly and GreenApps. Wi-Fly is a new connectivity paradigm designed for regions without Internet coverage that enables communication between a lightweight Wi-Fi device on commercial planes and ground stations. Through empirical experiments with test flights and simulation, we show that Wi-Fly and its extensions have the potential to provide connectivity in the most remote regions of the world. In GreenApps, we look at how localized cellular applications can be built for rural communities on top of software-defined cellular base stations. We deployed the GreenApps platform on rural base stations for communities in Ghana and Nicaragua and supported multiple localized applications for rural communities.

  Enhancing transport layer performance over cellular networks is critical to improve end-to-end application performance for mobile users. Cellular networks have unique challenges that make conventional transport protocols not suitable for these environments. In the past few years, several new delay-based congestion-control algorithms have been developed with complex nonlinear control loops for cellular contexts. While these protocols have shown promise, it has been extremely challenging to analyze and interpret the behavior of these algorithms especially under highly variable network conditions (e.g., cellular links). In the Model-Driven Interpretable (MDI) congestion control work, we provide a model-driven framework to reason about the behavior of such congestion control algorithms. Our modeling approach simplifies a congestion control algorithm’s behavior into a guided random walk over a two-dimensional Markov model. We show that the model of a congestion-control algorithm can give key insights into its convergence and performance. More recently, we also looked at how to learn early signals of congestion in highly varying 5G channels. In particular we worked with Wi-Gig traces collected at 60 GHz and showed that it is possible to learn highly accurate early congestion signals using delay features observed at end-hosts.

• Ph.D. Thesis 2019 Co-Located Augmented and Virtual Reality Systems DeFanti, Connor Abstract | PDF

  Title: Co-Located Augmented and Virtual Reality Systems

  Candidate: DeFanti, Connor

  Advisor(s): Perlin, Ken

  Abstract:

  Augmented and Virtual Reality (AVR) systems have become increasingly popular in the worlds of entertainment and industry. However, many current systems are limited in scope to experiences that isolate a single user within a given physical space. While many such experiences allow for interactions between remotely located users, very few experiences allow for multiple users to coexist in the same physical space while interacting with a consistent world-view of shared virtual objects. Our research has found that by enabling this co-located paradigm, users are able to have rich interactions that are otherwise impossible. This thesis presents a series of experiments that demonstrate the importance of the social aspects of co-located AVR, a set of solutions that overcome the difficulties often encountered in such experiences, and directions for future scalability using forthcoming hardware and technologies.

• Ph.D. Thesis 2019 Design for Customized Manufacturing Gil-Ureta, Francisca T. Abstract | PDF

  Title: Design for Customized Manufacturing

  Candidate: Gil-Ureta, Francisca T.

  Advisor(s): Denis Zorin

  Abstract:

  Over the past few years, 3D printing technology has captivated business and consumers alike with its promise of affordable custom manufacturing. The expectation is, in the future, people will be able to easily customize and manufacture objects to fit individual needs. To make this a reality, we need new methods that support the creative process of makers, from conception to fabrication.
  
  In this thesis, I present three projects where we reexamine the tools and workflows used for customized design. The core idea behind these projects is that, compared with traditional methods, we design for an unknown or changeable manufacturing process, which affects the life-cycles of design. Our goal is to create tools that simplify the modification, optimization, and evaluation of designs such that they can be easily altered to fit manufacturing and personal constraints.
  
  Although fabrication constraints are unlimited, we can study specific domains to learn the most common ones. In the first project, we present an interactive modeling tool for designing mechanical objects, which are determined mostly by kinematic constraints. In the second project, we study the structural efficiency of shells and introduce an efficient method for designing shell reinforcements of minimal weight. Finally, in the third project, we develop a robust collision resolution algorithm, crucial for the design and optimization of
  models subject to dynamic impulses.
  

• Ph.D. Thesis 2019 Scalable Machine Learning using Dataflow Graph Analysis Huang, Chien-Chin Abstract | PDF

  Title: Scalable Machine Learning using Dataflow Graph Analysis

  Candidate: Huang, Chien-Chin

  Advisor(s): Li, Jinyang

  Abstract:

  In the past decade, the abundance of computing resources and the growth of data boost the development of machine learning applications. Many computation frameworks, e.g., Hadoop, Spark, TensorFlow, and PyTorch, have been proposed and become widely used in the industry. However, programming large-scale machine learning applications is still challenging and requires the manual efforts of developers to achieve good performance.

  For example, when parallelizing arrays to hundreds of CPU machines, it is critical to choose a good partition strategy to co-locate the computation arrays to reduce network communication. Unfortunately, existing distributed array frameworks usually use a default partition scheme and requires manually partitioning if another parallel strategy is used, making it less easy to develop a distributed array program. Another example is running deep learning applications with GPU. Modern GPU can be orders of magnitude faster than CPU and becomes an attractive computation resource. Unfortunately, the limited memory size of GPU restricts the scale of the DNN models can be run. It is desired to have a computation framework to allow users to explore deeper and wider DNN models.

  Modern distributed frameworks generally adopt a dataflow-style programming paradigm. The dataflow graph of an application exposes valuable information to optimize the application. In this thesis, we present two techniques to address the above issues via dataflow graph analysis.

  We first design Spartan to help users parallelize distributed arrays on a CPU cluster. Spartan is a distributed array framework, built on top of a set of higher-order dataflow operators. Based on the operators, Spartan provides a collection of Numpy-like array APIs. Developers can choose the built-in array APIs or directly use the operators to construct machine learning applications. To achieve good performance for the distributed application, Spartan analyzes the communication pattern of the dataflow graph captured through the operators and applies a greedy strategy to find a good partition scheme to minimize the communication cost.

  To support memory-intensive deep learning applications on a single GPU, we develop SwapAdvisor, a swapping system that automatically swaps temporarily unused tensors from GPU memory to CPU memory. To minimize the communication overhead, SwapAdvisor analyzes the dataflow graph of the given DNN model and uses a custom-designed genetic algorithm to optimize the operator scheduling and memory allocation. Based on the optimized operator schedule and memory allocation, SwapAdvisor can determine what and when to swap to achieve a good performance.

• Ph.D. Thesis 2019 Compositional Abstractions for Verifying Concurrent Data Structures Krishna, Siddharth Abstract | PDF

  Title: Compositional Abstractions for Verifying Concurrent Data Structures

  Candidate: Krishna, Siddharth

  Advisor(s): Thomas Wies

  Abstract:

  Formal verification has had great success in improving the reliability of real-world software, with projects such as ASTREE, CompCert, and Infer showing that rigorous mathematical analysis can handle the scale of today's cyber-infrastructure. However, despite these successes, many core software components are yet to be verified formally. Concurrent data structures are a class of algorithms that are becoming ubiquitous, as software systems seek to make use of the increasingly parallel design of computers and servers. These data structures use sophisticated algorithms to perform fine-grained synchronization between threads, making them notoriously difficult to design correctly, with bugs being found both in actual implementations and in the designs proposed by experts in peer-reviewed publications. The rapid development and deployment of these concurrent algorithms has resulted in a rift between the algorithms that can be verified by the state-of-the-art techniques and those being developed and used today. The goal of this dissertation is to bridge this gap and bring the certified safety of formal verification to the concurrent data structures used in practice.

  Permission-based program logics such as separation logic have been established as the standard technique for verifying programs that manipulate complex heap-based data structures. These logics build on so-called separation algebras, which allow expressing properties of heap regions such that modifications to a region do not invalidate properties stated about the remainder of the heap. This concept is key to enabling modular reasoning and also extends to concurrency. However, certain data structure idioms prevalent in real-world programs, especially concurrent programs, are notoriously difficult to reason about, even in these advanced logics (e.g., random access into inductively defined structures, data structure overlays). The underlying issue is that while heaps are naturally related to mathematical graphs, many ubiquitous graph properties are non-local in character. Examples of such properties include reachability between nodes, path lengths, acyclicity and other structural invariants, as well as data invariants which combine with these notions. Reasoning modularly about such global graph properties remains a hard problem, since a local modification can have side-effects on a global property that cannot be easily confined to a small region.

  This dissertation addresses the question: What separation algebra can be used to prove that a program maintains a global graph property by reasoning only about the local region modified by the program? We propose a general class of global graph properties, that we call flows, that can be expressed as fixpoints of algebraic equations over graphs. Flows can encode structural properties of the heap (e.g. the reachable nodes from the root form a tree), data invariants (e.g. sortedness), as well as combinations of both shape and data constraints of overlaid structures in a uniform manner. We then introduce the notion of a flow interface, an abstraction of a region in the heap, which expresses the constraints and guarantees between the region and its context with respect to the flow. Under a suitable notion of composition that preserves the flow values, we show that flow interfaces form the desired separation algebra.

  Building on our theory of flows, we develop the flow framework, a general proof technique for modular reasoning about global graph properties over program heaps that can be integrated with existing separation logics. We further devise a strategy for automating this technique using SMT-based verification tools. We have implemented this strategy on top of the verification tool Viper and applied it successfully to a variety of challenging benchmarks including 1) algorithms involving general graphs such as Dijkstra's algorithm and a priority inheritance protocol, 2) inductive data structures such as linked lists and B trees, 3) overlaid data structures such as the Harris list and threaded trees, and 4) OO design patterns such as Composite and Subject/Observer. We are not aware of any single other approach that can handle these examples with the same degree of simplicity or automation.

  While the flow framework is applicable to any data structure, its features give rise to a new form of modular reasoning for certain concurrent data structures. Concurrent separation logics already apply modularity on multiple levels to simplify correctness proofs, decomposing them according to program structure, program state, and individual threads. Despite these advances, it remains difficult to achieve proof reuse across different data structure implementations. For the large class of concurrent search structures, we demonstrate how one can achieve further proof modularity by decoupling the proof of thread safety from the proof of structural integrity. We base our work on the template algorithms of Shasha and Goodman that dictate how threads interact but abstract from the concrete layout of nodes in memory. By using the flow framework of compositional abstractions in the separation logic Iris, we show how to prove correctness of template algorithms, and how to instantiate them to obtain multiple verified implementations. We demonstrate our approach by formalizing three concurrent search structure templates, based on link, give-up, and lock-coupling synchronization, and deriving implementations based on B-trees, hash tables, and linked lists. These case studies represent algorithms used in real-world file systems and databases, which have so far been beyond the capability of automated or mechanized state-of-the-art verification techniques. Our verification is split between the Coq proof assistant and the deductive verification tool GRASShopper in order to demonstrate that our proof technique and framework can be applied both in fully mechanized proof assistants as well as automated program verifiers. In addition, our approach reduces proof complexity and is able to achieve significant proof reuse.

• Ph.D. Thesis 2019 Parallel Contact-Aware Algorithms for Large-Scale Direct Blood Flow Simulations Lu, Libin Abstract | PDF

  Title: Parallel Contact-Aware Algorithms for Large-Scale Direct Blood Flow Simulations

  Candidate: Lu, Libin

  Advisor(s): Zorin, Denis

  Abstract:

  Experimental and theoretical evidence suggests that blood flow can be well approximated by a mixture model of a Newtonian fluid and deformable particles representing the red blood cells. We use a well-established boundary integral formulation for the problem as the foundation of our approach. This type of formulations, with a high-order spatial discretization and an implicit and adaptive time discretization, have been shown to be able to handle complex interactions between particles with high accuracy. Yet, for dense suspensions, very small time-steps or expensive implicit solves as well as a large number of discretization points are required to avoid non-physical contact and intersections between particles, lead- ing to infinite forces and numerical instability. Given the importance of vesicle flows, in this thesis we focus in efficient numerical methods for such problems: we present computationally parallel-scalable algorithms for the simulation of dense deformable vesicles in two and three dimensions both in unbounded and bounded domain.

  Our method maintains the accuracy of previous methods at a significantly lower cost for dense suspensions and the time step size is independent from the volume fraction. The key idea is to ensure interference-free configuration by introducing explicit contact constraints into the system. While such constraints are unnecessary in the formulation, in the discrete form of the problem, they make it possible to eliminate catastrophic loss of accuracy by preventing contact explicitly. Experimental and theoretical evidence suggests that blood flow can be well approximated by a mixture model of a Newtonian fluid and deformable particles representing the red blood cells. We use a well-established boundary integral formulation for the problem as the foundation of our approach. This type of formulations, with a high-order spatial discretization and an implicit and adaptive time discretization, have been shown to be able to handle complex interactions between particles with high accuracy. Yet, for dense suspensions, very small time-steps or expensive implicit solves as well as a large number of discretization points are required to avoid non-physical contact and intersections between particles, lead- ing to infinite forces and numerical instability. Given the importance of vesicle flows, in this thesis we focus in efficient numerical methods for such problems: we present computationally parallel-scalable algorithms for the simulation of dense deformable vesicles in two and three dimensions both in unbounded and bounded domain.

  Our method maintains the accuracy of previous methods at a significantly lower cost for dense suspensions and the time step size is independent from the volume fraction. The key idea is to ensure interference-free configuration by introducing explicit contact constraints into the system. While such constraints are unnecessary in the formulation, in the discrete form of the problem, they make it possible to eliminate catastrophic loss of accuracy by preventing contact explicitly.

  Introducing contact constraints results in a significant increase in stable time- step size for locally-implicit time-stepping, and a reduction in the number of points adequate for stability. Our method permits simulations with high volume fractions; we report results with up to 60% volume fraction. We demonstrated the parallel v scaling of the algorithms on up to 35K CPU cores.

• Ph.D. Thesis 2019 Leveraging Program Analysis for Type Inference Pavlinovic, Zvonimir Abstract | PDF

  Title: Leveraging Program Analysis for Type Inference

  Candidate: Pavlinovic, Zvonimir

  Advisor(s): Wies, Thomas

  Abstract:

  Type inference is a popular feature of programming languages used to automatically guarantee the absence of certain execution errors in programs at compile time. The convenience of type inference, unfortunately, comes with a cost. Developing type inference algorithms is a challenging task that currently lacks a systematic approach. Moreover, programmers often have problems interpreting error reports produced by type inference. The overarching goal of this thesis is to provide a mathematically rigorous framework for the systematic development of sophisticated type inference algorithms that are convenient to use by the programmers. To this end, we focus on two specific problems in this thesis: (1) how to constructively design type inference algorithms that improve over the state-of-the-art and (2) how to automatically debug type errors that arise during inference. We base our approach on the observation that, similar to type inference, program analysis algorithms automatically discover various program properties that can be used to show program correctness. Type inference and program analysis techniques, although similar, have traditionally been developed independently of each other. In contrast, this thesis further explores the recent path of leveraging program analysis for type inference.

  As our first contribution, we use abstract interpretation to constructively design type inference algorithms. We specifically focus on Liquid types, an advanced family of algorithms that combine classical typing disciplines and known static analyses to prove various safety properties of functional programs. By using abstract interpretation, we make the design space of Liquid type inference explicit. We also unveil the general type inference framework underlying Liquid types. By properly instantiating this general framework, one obtains novel type inference algorithms that are sound by construction.

  Our second contribution is a framework for automatically debugging type errors for languages that deploy type inference in the style of Hindley-Milner, such as OCaml and Haskell. Such languages are notorious for producing cryptic type error reports that are often not helpful in fixing the actual bug. We formulate the problem of finding the root cause of type errors as an optimization problem expressed in a formal logic. We then show how to solve this problem using automated theorem provers. We experimentally illustrate how our framework can efficiently produce type error reports that outperform the state-of-the-art solutions in identifying the true cause of type errors.

  In summary, this thesis introduces a mathematical framework for the systematic design of sophisticated type inference algorithms that are sound by construction. Our results further enable automatic generation of more meaningful type error diagnostics, ultimately making type inference more usable by the programmers.

• Ph.D. Thesis 2019 Concentration and Anti-concentration for Markov Chains Rao, Shravas Abstract | PDF

  Title: Concentration and Anti-concentration for Markov Chains

  Candidate: Rao, Shravas

  Advisor(s): Regev, Oded

  Abstract:

  We study tail bounds and small ball probabilities for sums of random variables obtained from a Markov chain. In particular, we consider the following sum \(S_n = f_1(Y_1)+\cdots+f_n(Y_n)\) where \(\{Y_i\}_{i=1}^{\infty}\) is a Markov chain with state space \([N]\), transition matrix \(A\), and stationary distribution \(\mu\) such that \(Y_1\) is distributed as \(\mu\), and \(f_i: [N] \rightarrow \mathbb{R}\). We also consider settings in which \(f_i(Y_i)\) is vector-valued.

  In all results, the bounds are in terms of the spectral gap of the Markov chain. In almost all of the results in this thesis, when the transitions are independent and the spectral gap is \(1\), the bounds match the corresponding bounds for independent random variables up to constant factors.

  We first obtain tail bounds in the case that only the \(p\)th moment of the random variable \(f_i(Y_i)\) is bounded. This is a Markov chain version of a corollary of the Marcinkiewicz–Zygmund inequality. Using this, we also obtain tail bounds for \(S_n\) when the \(f_i(Y_i)\) are elements of an \(\ell_q\) space.

  Next, we obtain sharp tail bounds when the random variables \(f_i(Y_i)\) are bounded and the expected value of \(S_n\) is small. This is a Markov chain version of a Poisson approximation to sums of independent random variables. As an application, we explain how such tail bounds can be used to construct simple and explicit resilient functions that match the non-constructive functions shown to exist due to the work of Ajtai and Linial.

  Next, we obtain tail bounds in the case that the \(f_i(Y_i)\) are bounded in the range \([-a_i, a_i]\) for each \(i\). This is a Markov chain version of the Hoeffding inequality. This improves upon previously known bounds in that the dependence is on \(\sqrt{a_1^2+\cdots+a_n^2}\) rather than \(\max_{i}\{a_i\}\sqrt{n}.\) Using this, we obtain tail bounds for certain types of random variables in which the \(f_i(Y_i)\) are elements of any Banach space.

  Finally, we show that if the \(f_i(Y_i)\) take on values \(\{-a_i, a_i\}\) with equal probability and the \(a_i\) are Euclidean vectors with norm at least \(1\), the probability that \(S_n\) lies in a ball of volume \(1\) is small. This is a Markov chain version of the Littlewood-Offord inequality. We also construct a new pseudorandom generator for the Littlewood-Offord problem.

• Ph.D. Thesis 2019 Approximation algorithms, Hardness, and PCPs Thiruvenkatachari, Devanathan Abstract | PDF

  Title: Approximation algorithms, Hardness, and PCPs

  Candidate: Thiruvenkatachari, Devanathan

  Advisor(s): Khot, Subhash

  Abstract:

  This thesis is a collection of theoretical results on the topic of approximation algorithms and hardness of approximation. The results presented here use a combination of classical and modern techniques to achieve better approximation algorithms and hardness results for some pivotal NP-hard problems and their variants. We study CSPs from a multi-objective point of view, with the goal of simultaneous optimization of multiple instances over the same set of variables, with MAX-CUT as the central focus. We provide an approximation algorithm that is near optimal assuming the unique games conjecture. We also study PCPs and their role in hardness of approximation, and present a hardness result for 3-LIN in the sub-constant soundness regime. Lastly, dictatorship testing is a property testing problem with significant applications in proving hardness results, and we present an improvement on the soundness of the k-bit dictatorship test with perfect completeness.

• Ph.D. Thesis 2019 Tactile Perception Design for Fabrication Tymms, Chelsea Abstract | PDF

  Title: Tactile Perception Design for Fabrication

  Candidate: Tymms, Chelsea

  Advisor(s): Zorin, Denis

  Abstract:

  High-resolution 3D printing technology provides the ability to manufacture shapes with precise geometry. Controlling this fine-scale geometry to confer haptic qualities is a growing area of research in fabrication. In this thesis, I will present three projects addressing the question of how to fabricate surface textures with controlled tactile properties and exploring how tactile textures can be used in custom manufacturing and to expand the understanding of the human sense of touch.

  Surface roughness is one of the most significant qualities in haptic perception, essential to material identification, comfort, and usability. Past perceptual studies on roughness have typically used stimuli that are existing materials or in a narrow range of custom-made materials. In the first project presented in this thesis, we explore the use of 3D printing to manufacture stimuli. We used modeling and 3D printing to manufacture a set of fine parametric bump textures, and we used these texture stimuli in a psychophysical study of human roughness perception. We investigated the contribution of the texton spacing, size, and arrangement to the texture's perceived tactile roughness.

  In the second project, we quantitatively address the problem of mapping arbitrary texture geometry to tactile roughness. Drawing from insights in past neurophysiology research, we developed a model that simulates human touch to predict a texture's tactile roughness from its surface geometry. We fabricated a set of 46 parametric and real-life textures, and we used psychophysical experiments with human subjects to place them in the perceptual space for tactile roughness using non-metric multidimensional scaling. We closely match this space with our quantitative model, obtained from strain fields derived from the elasticity simulations of the human skin contacting texture geometry. We demonstrate how this model can be applied to predict and alter surface roughness, and we show several applications in the context of fabrication.

  The third project extends these ideas by developing a method to control a texture's haptic qualities and visual appearance at the same time. The tactile feeling and visual appearance of objects often interact in unpredictable ways, and both serve important purposes for identification and usability. In this project, we develop an optimization method to maintain a texture's visual appearance while altering its perceived tactile roughness or tactile temperature. Our optimization method, which is enabled by neural network-based models, allows us to change a texture to a different desired tactile feeling while preserving the visual appearance, at a relatively low computational cost.

• Ph.D. Thesis 2019 End-to-End Learning for Autonomous Driving Zhang, Jiakai Abstract | PDF

  Title: End-to-End Learning for Autonomous Driving

  Candidate: Zhang, Jiakai

  Advisor(s): Cho, Kyunghyun

  Abstract:

  The end-to-end learning approach for autonomous driving has sparked great interest in both academic and industry in recent years. The approach can be defined as learning a model that maps from sensory input, such as image frames from a camera, to driving actions for controlling the autonomous vehicle such as steering. Compared to the traditional autonomous driving system, which often includes perception, localization, mapping, and path planning, the end-to-end learning approach offers a more efficient method of utilizing large amounts of expert driver demonstrations to achieve fully autonomous driving without acquiring expensive labeled data such as bounding box for objects.

  The end-to-end learning for autonomous driving can be done by supervised learning, where a model is tuned to minimize the difference between predicted actions and ground-truth actions. The ground truth of driving actions is usually obtained from driver demonstrations. A model trained in this way, however, suffers from unexpected behaviors due to the mismatch between the samples visited by a learned model and the samples collected by an expert driver. To address this issue, we first introduce an end-to-end supervised learning approach with data augmentation to train a model to keep a vehicle driving at the center of a lane. The data augmentation is done by synthetically generating new samples through rotating and translating input images captured from a front-facing camera and calculating compensatory steering. We show that using such automatically-augmented data, a trained model can drive a car to follow a lane in various conditions on highways and local and residential roads.

  Instead of generating augmented data, we can also collect new samples when trying out the learned model. Aiming to reduce the number of times querying an expert for labeling, we propose SafeDAgger algorithm, which is a query-efficient imitation learning approach. We show that our method significantly reduces the number of querying times and trains a driving model more efficiently. A model trained by our proposed SafeDAgger algorithm can successfully drive a racing car in a simulator to do lane following and overtaking.

  The expert demonstrations provided by humans and used for training models often show significant variability due to latent factors. Given such expert demonstrations, a model trained by minimizing the difference between the expert driving actions and predicted driving actions can output dangerous driving actions that may cause serious accidents. We address this issue by introducing a variational mixture density network to model the variability using a discrete latent variable. The experimental results in a racing car simulator show that the model trained using our proposed method can learn the variability of driving signals from expert demonstrations and successfully distinguish certain driving behaviors such as changing lanes and following lanes.

  We introduce a simulator to support the development, training, and evaluation of autonomous driving systems using the end-to-end learning approaches. Leveraging this simulator, we demonstrate how to train and evaluate models to drive a truck that follows a navigation map in a video game.

  In summary, this thesis introduces the end-to-end learning approaches for autonomous driving to address the data mismatch issue and learn the variability of expert driving actions. Our results show that the trained model can drive the vehicle to follow a lane, change lanes and make turns in simulated driving environments.

• Ph.D. Thesis 2019 Text Representation using Convolutional Networks Zhang, Xiang Abstract | PDF

  Title: Text Representation using Convolutional Networks

  Candidate: Zhang, Xiang

  Advisor(s): LeCun, Yann

  Abstract:

  This dissertation applies convolutional networks for learning representations of text, and it consists of several parts. The first part offers an empirical exploration on the use of character-level convolutional networks (ConvNets) for text classification. We constructed several large-scale datasets to show that character-level convolutional networks could achieve state-of-the-art or competitive results. Comparisons are offered against traditional models such as bag of words, n-grams and their TFIDF variants, and deep learning models such as word-based ConvNets and recurrent neural networks. These results indicate that using low-level inputs – in this case characters – for convolutional networks could be feasible for text representation learning.

  The second part concerns which text encoding method might work for convolutional networks. We include a comprehensive comparison of different encoding methods for the task of text classification using 14 large-scale datasets in 4 languages including Chinese, English, Japanese and Korean. Different encoding levels are studied, including UTF-8 bytes, characters, words, romanized characters and romanized words. For all encoding levels, whenever applicable, we provide comparisons with linear models, fastText and convolutional networks. For convolutional networks, we compare between encoding mechanisms using character glyph images, one-hot (or one-of-n) encoding, and embedding. From these 473 models, one of the conclusions is that byte-level one-hot encoding works consistently best for convolutional networks.

  Based on this, in the third part of the dissertation we develop a convolutional network at the level of bytes for learning representations through the task of auto-encoding. The proposed model is a multi-stage deep convolutional encoder-decoder framework using residual connections, containing up to 160 parameterized layers. Each encoder or decoder contains a shared group of modules that consists of either pooling or up-sampling layers, making the network recursive in terms of abstraction levels in representation. The decoding process is non-sequential. Results for 6 large-scale paragraph datasets are reported, in 3 languages including Arabic, Chinese and English. Analyses are conducted to study several properties of the proposed model. Experiments are presented to verify that the auto-encoder can learn useful representations.

  In the fourth part of the dissertation, we use the improved design from the previous auto-encoding model to text classification, adding comparisons between residual and dense connections. This further validates the choice of the architecture we made for the auto-encoding model, and the effectiveness of the recursive architecture with residual or dense connections.

• Ph.D. Thesis 2019 Unsupervised Learning with Regularized Autoencoders Zhao, Junbo Abstract | PDF

  Title: Unsupervised Learning with Regularized Autoencoders

  Candidate: Zhao, Junbo

  Advisor(s): Yann LeCun

  Abstract:

  Deep learning has enjoyed remarkable successes in a variety of domains.These successes often emerge at the cost of large annotated datasets and training computationally heavy neural network models.The learning paradigm for this is called \emph{supervised learning}. However, to reduce the sample complexity while improving the universality of the trained models is a crucial next step that may to artificial intelligence. \emph{Unsupervised Learning}, in contrast to supervised learning, aims to build neural network models with more generic loss objectives requiring little or no labelling effort, and therefore it does not reside on any specific domain-task. In spite of the brevity of its goal, unsupervised learning is a broad topic that relates or includes several sub-fields, such as density estimation, generative modeling, world model and etc. In this thesis, we primarily adopt an energy-based view unifying these different fields~\citep{lecun2006tutorial}. A desired energy function reflects the data manifold by differentiating the energy assigned to the points on the data manifold against points off the manifold. With this foundation, we first cast the popular autoencoder and adversarial learning framework into an energy-based perspective, and then propose several technique or architectures with a motivation to learn better-shaped energy function. We also show that the proposed techniques in this thesis cover a wide spectrum of applications including image/text generative modeling, text summarization, style-transfer without aligned data, transfer/semi-supervised learning on both computer vision and natural language processing. The thesis is organized as follows. First, we assess the validity and the main challenges of energy-based learning. We then introduce two frameworks focusing on strengthening autoencoders by building unit connection hierarchies via either hard-coded pooling or self-learned graphs. Finally, we propose several systematic regularization techniques, based on adversarial training and vector discretization.

• Ph.D. Thesis 2019 Unsupervised Learning with Regularized Autoencoders Zhao, Junbo Abstract | PDF

  Title: Unsupervised Learning with Regularized Autoencoders

  Candidate: Zhao, Junbo

  Advisor(s): LeCun, Yann

  Abstract:

  Deep learning has enjoyed remarkable successes in a variety of domains. These successes often emerge at the cost of large annotated datasets and training computationally heavy neural network models. The learning paradigm for this is called supervised learning. However, to reduce the sample complexity while improving the universality of the trained models is a crucial next step that may to artificial intelligence. Unsupervised Learning, in contrast to supervised learning, aims to build neural network models with more generic loss objectives requiring little or no labelling effort, and therefore it does not reside on any specific domain-task. In spite of the brevity of its goal, unsupervised learning is a broad topic that relates or includes several sub-fields, such as density estimation, generative modeling, world model and etc. In this thesis, we primarily adopt an energy-based view unifying these different fields. A desired energy function reflects the data manifold by differentiating the energy assigned to the points on the data manifold against points off the manifold. With this foundation, we first cast the popular autoencoder and adversarial learning framework into an energy-based perspective, and then propose several technique or architectures with a motivation to learn better-shaped energy function. We also show that the proposed techniques in this thesis cover a wide spectrum of applications including image/text generative modeling, text summarization, style-transfer without aligned data, transfer/semi-supervised learning on both computer vision and natural language processing. The thesis is organized as follows. First, we assess the validity and the main challenges of energy-based learning. We then introduce two frameworks focusing on strengthening autoencoders by building unit connection hierarchies via either hard-coded pooling or self-learned graphs. Finally, we propose several systematic regularization techniques, based on adversarial training and vector discretization.

• Ph.D. Thesis 2018 Deep Generative Models of Images and Video Denton, Emily Lynn Abstract | PDF

  Title: Deep Generative Models of Images and Video

  Candidate: Denton, Emily Lynn

  Advisor(s): Fergus, Rob

  Abstract:

  Deep neural networks have seen wide success in the supervised setting in recent years. Many of these successes rely heavily on large training sets of manually annotated data. Given the difficulty of obtaining enough labeled data to scale many deep learning approaches, it is increasingly important to look for better methods of utilizing large amounts of unlabeled data. Building generative models of images and video is a fundamental paradigm of learning from unlabeled data. Unsupervised criterion based on generating or reconstructing images drive many representation learning frameworks. Video is a particularly appealing domain for unsupervised learning due to the inherent temporal structure of the data. This structure lends itself to representation learning approaches based on extracting invariances and predicting future frames, given the past.

  Additionally, building accurate models of the world that facilitate future prediction can be useful for model based reinforcement learning, planning, and more generally, endowing an agent with the capacity to reason about its environment. Incorporating predictive models can potentially help alleviate the sample inefficiency of many reinforcement learning systems.

  In this thesis, we review the challenges associated with generating images and videos. We then introduce a multi-scale image generation framework that demonstrates impressive performance on real world image datasets. This method was the first to demonstrate empirically the potential of generative adversarial networks. We also address two challenging aspects of video generation:learning a latent space that affords easier prediction and modeling the uncertainty in video sequences.

• Ph.D. Thesis 2018 Deep Networks for Forward Prediction and Planning Henaff, Mikael Bruce Abstract | PDF

  Title: Deep Networks for Forward Prediction and Planning

  Candidate: Henaff, Mikael Bruce

  Advisor(s): LeCun, Yann

  Abstract:

  Learning to predict how an environment will evolve and the consequences of one’s actions is an important ability for autonomous agents, and can enable planning with relatively few interactions with the environment which may be slow or costly. However, learning an accurate forward model is often difficult in practice due to several features often present in complex environments. First, many environments exhibit long-term dependencies which require the system to learn to record and maintain relevant information in its memory over long timescales. Second, the environment may only be partially observed, and the aspects of the environment which are observed may depend on parts of the environment which are hidden. Third, many observed processes contain some form of apparent or inherent stochasticity, which makes the task of predicting future states ill-defined. In this thesis, we propose approaches to tackle and better understand these different challenges associated with learning predictive models of the environment and using them for planning. We first provide an analysis of recurrent neural network (RNN) memory, which sheds light on the mechanisms by which RNNs are able to store different types of information in their memory over long timescales through the analysis of two synthetic benchmark tasks. We then introduce a new neural network architecture which keeps an estimate of the state of the environment in its memory, and can deal with partial observability by reasoning based on what is observed. We next present a new method for performing planning using a learned model of the environment with both discrete and continuous actions. Finally, we propose an approach for model-based planning in the presence of both environment uncertainty and model uncertainty, and evaluate it on a new real-world dataset and environment with applications to autonomous driving.

• Ph.D. Thesis 2018 Learning Representations of Text through Language and Discourse Modeling: From Characters to Sentences Jernite, Yacine Abstract | PDF

  Title: Learning Representations of Text through Language and Discourse Modeling: From Characters to Sentences

  Candidate: Jernite, Yacine

  Advisor(s): Sontag, David

  Abstract:

  In this thesis, we consider the problem of obtaining a representation of the meaning expressed in a text. How to do so correctly remains a largely open problem, combining a number of inter-related questions (e.g. what is the role of context in interpreting text? how should language understanding models handle compositionality? etc...) In this work, after reflecting on some of these questions and describing the most common sequence modeling paradigms in use in recent work, we focus on two specifically: what level of granularity text should be read at, and what training objectives can lead models to learn useful representations of a text’s meaning.

  In a first part, we argue for the use of sub-word information for that purpose, and present new neural network architectures which can either process words in a way that takes advantage of morphological information, or do away with word separations altogether while still being able to identify relevant units of meaning.

  The second part starts by arguing for the use of language modeling as a learning objective, and provides algorithms which can help with its scalability issues and propose a globally rather than locally normalized probability distribution. It then explores the question of what makes a good language learning objective, and introduces discriminative objectives inspired by the notion of discourse coherence which help learn a representation of the meaning of sentences.

• Ph.D. Thesis 2018 Deep Learning for Information Extraction Nguyen, Thien Huu Abstract | PDF

  Title: Deep Learning for Information Extraction

  Candidate: Nguyen, Thien Huu

  Advisor(s): Grishman, Ralph

  Abstract:

  The explosion of data has made it crucial to analyze the data and distill important information effectively and efficiently. A significant part of such data is presented in unstructured and free-text documents. This has prompted the development of the techniques for information extraction that allow computers to automatically extract structured information from the natural free-text data. Information extraction is a branch of natural language processing in artificial intelligence that has a wide range of applications, including question answering, knowledge base population, information retrieval etc. The traditional approach for information extraction has mainly involved hand-designing large feature sets (feature engineering) for different information extraction problems, i.e, entity mention detection, relation extraction, coreference resolution, event extraction, and entity linking. This approach is limited by the laborious and expensive effort required for feature engineering for different domains, and suffers from the unseen word/feature problem of natural languages.

  This dissertation explores a different approach for information extraction that uses deep learning to automate the representation learning process and generate more effective features. Deep learning is a subfield of machine learning that uses multiple layers of connections to reveal the underlying representations of data. I develop the fundamental deep learning models for information extraction problems and demonstrate their benefits through systematic experiments.

  First, I examine word embeddings, a general word representation that is produced by training a deep learning model on a large unlabelled dataset. I introduce methods to use word embeddings to obtain new features that generalize well across domains for relation extraction. This is done for both the feature-based method and the kernel-based method of relation extraction.

  Second, I investigate deep learning models for different problems, including entity mention detection, relation extraction and event detection. I develop new mechanisms and network architectures that allow deep learning to model the structures of information extraction problems more effectively. Some extensive experiments are conducted on the domain adaptation and transfer learning settings to highlight the generalization advantage of the deep learning models for information extraction.

  Finally, I investigate the joint frameworks to simultaneously solve several information extraction problems and benefit from the inter-dependencies among these problems. I design a novel memory augmented network for deep learning to properly exploit such inter-dependencies. I demonstrate the effectiveness of this network on two important problems of information extraction, i.e, event extraction and entity linking.

• Ph.D. Thesis 2018 Accelerating Approximate Simulation with Deep Learning Schlachter, Kristofer Abstract | PDF

  Title: Accelerating Approximate Simulation with Deep Learning

  Candidate: Schlachter, Kristofer

  Advisor(s): Perlin, Ken

  Abstract:

  Once a simulation resorts to an approximate numerical solution one is faced with various tradeoffs in accuracy versus computation time. We propose that another approximate solution can be learned for two chosen simulations, which in our case, are just as useful but can be made faster to compute. The two problems addressed in this thesis are fluid simulation and the simulation of diffuse inter-reflection in computer graphics.

  Real-time simulation of fluid and smoke is a long standing problem in computer graphics, where state-of-the-art approaches require large compute resources, making real-time applications often impractical. In this work, we propose a data-driven approach that leverages the approximation power of deep-learning methods with the precision of standard fluid solvers to obtain both fast and highly realistic simulations. The proposed method solves the incompressible Euler equations following the standard operator splitting method in which a large, often ill-condition linear system must be solved. We propose replacing this system by learning a Convolutional Network (ConvNet) from a training set of simulations using a semi-supervised learning method to minimize long-term velocity divergence.

  ConvNets are amenable to efficient GPU implementations and, unlike exact iterative solvers, have fixed computational complexity and latency. The proposed hybrid approach restricts the learning task to a linear projection without modeling the well understood advection and body forces. We present real-time 2D and 3D simulations of fluids and smoke; the obtained results are realistic and show good generalization properties to unseen geometry.

  The next simulation that we address is the synthesis of images for training convnets. A challenge with training deep learning models is that they commonly require a large corpus of training data and retrieving sufficient real world data may be unachievable. A solution to this problem can be found in the use of synthetic or simulated training data. However, for simulated photographs or renderings, there hasn't been a systematic approach to comparing the relative benefits of different techniques in image synthesis.

  We compare multiple synthesis techniques to one another as well as the real data that they seek to replicate. We also introduce learned synthesis techniques that either train models better than the most realistic graphical methods used by standard rendering packages or else approach their fidelity using far less computation. We accomplish this by learning shading of geometry as well as denoising the results of low sample Monte Carlo image synthesis. Our major contributions are (i) a dataset that allows comparison of real and synthetic versions of the same scene, (ii) an augmented data representation that boosts the stability of learning, and (iii) three different partially differentiable rendering techniques where lighting, denoising and shading are learned. Finally we are able to generate datasets that can outperform full global illumination rendering and approach the performance of training on real data.

• Ph.D. Thesis 2018 Elements of Intelligence: Memory, Communication and Intrinsic Motivation Sukhbaatar, Sainbayar Abstract | PDF

  Title: Elements of Intelligence: Memory, Communication and Intrinsic Motivation

  Candidate: Sukhbaatar, Sainbayar

  Advisor(s): Fergus, Rob

  Abstract:

  Building an intelligent agent that can learn and adapt to its environment has always been a challenging task. This is because intelligence consists of many different elements such as recognition, memory, and planning. In recent years, deep learning has shown impressive results in recognition tasks. The aim of this thesis is to advance the deep learning techniques to other elements of intelligence.

  We start our investigation with memory, an integral part of intelligence that bridges past experience with current decision making. In particular, we focus on the episodic memory, which is responsible for storing our past experiences and recalling them. An agent without such memory will struggle at many tasks such as having a coherent conversation. We show that a neural network with an external memory is better at such tasks, outperforming traditional recurrent networks with an internal memory.

  Another crucial ingredient of intelligence is the capability to communicate with others. In particular, communication is essential for cooperative tasks, enabling agents to better collaborate and improve their division of labor. We investigate whether agents can learn to communicate from scratch without any external supervision. Our finding is that communication through a continuous vector facilitates faster learning by allowing gradients to flow between agents.

  Lastly, an intelligent agent must have an intrinsic motivation to learn about its environment on its own without any external supervision or rewards. Our investigation led to one such learning strategy where an agent plays a two-role game with itself. The first role proposes a task, and the second role tries to execute it. Since their goal is to make the other fail, their adversarial interplay pushes them to explore increasingly complex tasks, which results in a better understanding of the environment.

• Ph.D. Thesis 2018 Rethinking Customer Segmentation and Demand Learning in the Presence of Sparse, Diverse, and Large-scale Data Venkataraman, Ashwin Abstract | PDF

  Title: Rethinking Customer Segmentation and Demand Learning in the Presence of Sparse, Diverse, and Large-scale Data

  Candidate: Venkataraman, Ashwin

  Advisor(s): Jagabathula, Srikanth; Subramanian, Lakshminarayanan

  Abstract:

  Firms are now able to collect unprecedented amounts of data. This wealth of data provides new opportunities and capabilities for the firm to better solve classical problems within operational and marketing contexts, such as customer segmentation and demand learning. At the same time, the data imposes new challenges. In addition to its large-scale nature which creates computational issues, the data comes from a diversity of sources, varying in their respective measurement scales (e.g., clicks, ratings, purchase signals, etc.), and is typically sparse, containing a large fraction of missing observations. The diversity in the data makes it hard to directly compare different observations (clicks vs purchases, for instance) and the severe sparsity precludes any meaningful imputations of unobserved entries. The data also comes from unreliable sources, which introduce both unintentional and deliberate errors. The identities of such sources is very often unknown, which makes it difficult to determine which sources to trust.
  
  These data challenges require a rethink of traditional techniques for customer segmentation and demand learning. Given their importance and widespread use, this dissertation revisits the classical problems of customer segmentation and demand learning but in the presence of sparse, diverse, and large-scale data. The key contribution of the dissertation is a suite of novel methodologies to deal with the challenges described above.
  
  Part I of the dissertation focuses on the problem of customer segmentation. In Chapter 1, we consider the problem of segmenting (or clustering) a large population of customers based on their preferences, when the preference signals (e.g., clicks, ratings, etc.) come from a multitude of diverse data sources and each customer provides only a few observations. These data characteristics preclude the applicability of traditional marketing techniques as well as standard clustering approaches in machine learning. We propose a model-based embedding technique which takes the customer observations and a probabilistic model class generating the observations as inputs, and outputs an embedding—a low-dimensional vector representation in Euclidean space—for each customer. We then cluster the embeddings to obtain the segments. We show that our segmentation technique can be used to generate highly accurate personalized recommendations in two real-world case studies, including up to 8% improvement over the existing approach on an eBay dataset consisting of millions of customers and items. In addition, it outperforms (both in speed and accuracy) standard techniques in marketing and machine learning.
  
  In Chapter 2, we turn our attention to the domain of crowdsourced labeling, which provides a low-cost, easy and scalable way to collect labels from the crowd—composed of "workers"—which are then aggregated and used as inputs for training machine learning applications. The main challenge is that workers are often unreliable, and therefore can introduce unintentional or even intentional errors into the labels. The reliabilities of the workers are a priori unknown, so correctly aggregating the labels becomes difficult. We propose algorithms to separate the worker population into two segments, what we call "honest" and "adversarial" workers. Honest workers can provide incorrect labels, but their errors are probabilistic and therefore, can be corrected. Adversarial workers, on the other hand, adopt arbitrary labeling strategies (whether deterministic or probabilistic) and therefore, their labels cannot be trusted. We demonstrate that discarding the labels provided by even a few adversarial workers can significantly improve the accuracy of several existing approaches for aggregating the labels in real-world crowdsourcing datasets.
  
  Part II is devoted to demand learning. In Chapter 3, we consider the problem of learning customer demand for a set of substitutable products. Within operations, the customer demand is typically modeled using a mixture of logit models, which can capture heterogeneity as well as rich substitution patterns in customer preferences. The mixture model is fit to historical sales transactions and inventory data, and the fitted model is used to inform pricing and assortment decisions. We propose a novel nonparametric estimator for the mixture of logit models, providing the ability to make effective use of the large amounts of transaction data that firms have access to. By contrast, most existing techniques impose parametric assumptions—usually driven by tractability considerations—on the mixing distribution, and consequently can suffer from model misspecification issues. We show that our estimator is able to recover good approximations of different ground-truth mixing distributions—despite having no knowledge of their underlying structure—and outperforms the standard expectation-maximization (EM) benchmark in predictive and decision accuracies, while being an order of magnitude faster.

• Ph.D. Thesis 2017 On Quadtrees, Voronoi Diagrams, and Lattices: Results in Geometric Algorithms Bennett, Huxley Abstract | PDF

  Title: On Quadtrees, Voronoi Diagrams, and Lattices: Results in Geometric Algorithms

  Candidate: Bennett, Huxley

  Advisor(s): Yap, Chee

  Abstract:

  We present several results on geometric algorithms, and somewhat more specifically on algorithmic aspects of geometric structures including quadtrees, Voronoi diagrams, and lattices. Our work contains two parts, the first of which is on subdivision algorithms, and the second of which is on lattice algorithms.

  Subdivision algorithms amount to recursively splitting an ambient space into smaller pieces until certain conditions hold. Often the underlying space is a square in the plane (or a box in higher dimensions), whose subdivision is represented by a quadtree (or its higher-dimensional analogs). A quadtree is smooth if any two adjacent leaf boxes differ by at most one in depth. We first study the cost of the smooth split operation in quadtrees, showing that it has constant amortized cost in quadtrees of any fixed dimension.

  We then present a subdivision-based algorithm for computing isotopic epsilon-approximations of planar minimization diagrams. Given a family of continuous functions, its minimization diagram partitions the plane into regions on which each function is minimal. Minimization diagrams generalize many natural Voronoi diagrams, and we show how to use our framework to compute an anisotropic Voronoi diagram on polygonal sites. We have implemented a prototype of our algorithm for anisotropic Voronoi diagrams, and we provide experimental results.

  We then turn to studying lattice algorithms. A lattice is a regular ordering of points in Euclidean space, which is represented as the set of all integer combinations of some linearly independent vectors (which we call a basis of the lattice). In our first work on lattices, we introduce and study the Lattice Distortion Problem (LDP). LDP asks how "similar" two lattices are, i.e., what the minimum distortion of a linear bijection between two lattices is. We show how to compute low-distortion mappings with a tradeoff between approximation quality and running time based on a notion of basis reduction introduced by Seysen (Combinatorica 1993). We also show that LDP is NP-hard to approximate to within any constant factor (under randomized reductions).

  Finally, we study the problem of finding lattice bases which are optimal with respect to two basis quality measures. Namely, we study the problem of finding bases with minimal orthogonality defect, and with nearly minimal Seysen condition number. We give algorithms which solve both problems while running in time depending only on the rank of the lattice times a polynomial in the input length.

• Ph.D. Thesis 2017 Improving Event Extraction: Casting a Wider Net Cao, Kai Abstract | PDF

  Title: Improving Event Extraction: Casting a Wider Net

  Candidate: Cao, Kai

  Advisor(s): Grishman, Ralph

  Abstract:

  Information extraction is the task of automatically extracting structured information from unstructured and/or semi-structured machine-readable documents. One facet of information extraction is event extraction (EE): identifying instances of selected types of events appearing in natural language text. For each instance, EE should identify the type of the event, the event trigger (the word or phrase which evokes the event), the participants in the event, and (where possible) the time and place of the event.

  One EE task was defined and intensively studied as part of the ACE (Automatic Content Extraction) research program. The 2005 ACE EE task involved 8 types and 33 subtypes of events. For instance, given the sentence "She was killed by an automobile yesterday.", an EE system should be able to recognize the word "killed" as a trigger for an event of subtype DIE, and discover "an automobile" and "yesterday" as the Agent and Time arguments. This task is quite challenging, as the same event might appear in the form of various trigger expressions and an expression might represent different types of events in different contexts.

  To support the development and evaluation of ACE EE systems, the Linguistic Data Consortium annotated a text corpus (consisting primarily of news articles) with information on the events mentioned. This corpus was widely used to train ACE EE systems. However, the event instances in the ACE corpus are not evenly distributed, and so some frequent expressions involving ACE events do not appear in the training data, adversely affecting performance.

  The thesis presents several strategies for improving the performance of EE. We first demonstrate the effectiveness of two types of linguistic analysis -- dependency regularization and Abstract Meaning Representation -- in boosting EE performance. Next we show the benefit of an active learning strategy in which a person is asked to judge a limited number of phrases which may be event triggers. Finally we report the impact of combining our baseline system with event patterns from a system developed for a different EE task (the TABARI program). This step contains expert-level patterns generated by other research groups. Because the information received is complicated and quite different from the original corpus (ACE), the integration of this information requires more complex processing.

• Ph.D. Thesis 2017 Random Growth Models Florescu, Laura Abstract | PDF

  Title: Random Growth Models

  Candidate: Florescu, Laura

  Advisor(s): Spencer, Joel

  Abstract:

  This work explores variations of randomness in networks, and more specifically, how drastically the dynamics and structure of a network change when a little bit of information is added to "chaos". On one hand, I investigate how much determinism in diffusions de-randomizes the process, and on the other hand, I look at how superposing "planted" information on a random network changes its structure in such a way that the "planted" structure can be recovered.

  The first part of the dissertation is concerned with rotor-router walks, a deterministic counterpart to random walk, which is the mathematical model of a path consisting of a succession of random steps. I study and show results on the volume (``the range") of the territory explored by the random rotor-router model, confirming an old prediction of physicists.

  The second major part in the dissertation consists of two constrained diffusion problems. The questions in this model are to understand the long-term behavior of the models, as well as how the boundary of the processes evolves in time.

  The third part is detecting communities in, or more generally, clustering networks. This is a fundamental problem in mathematics, machine learning, biology and economics, both for its theoretical foundations as well as for its practical implications. This problem can be viewed as "planting" some structure in a random network; for example, in cryptography, a code can be viewed as hiding some integers in a random sequence. For such a model with two communities, I show both information theoretic thresholds when it is impossible to recover the communities based on the density of the edges "planted" between the communities, as well as thresholds for when it is computationally possible to recover the communities.

• Ph.D. Thesis 2017 Zero-knowledge Proofs: Efficient Techniques for Combination Statements and their Applications Ganesh, Chaya Abstract | PDF

  Title: Zero-knowledge Proofs: Efficient Techniques for Combination Statements and their Applications

  Candidate: Ganesh, Chaya

  Advisor(s): Dodis, Yevgeniy

  Abstract:

  Zero-knowledge proofs provide a powerful tool, which allows a prover to convince a verifier that a statement is true without revealing any further information. It is known that every language in NP has a zero knowledge proof system, thus opening up several cryptographic applications. While true in theory, designing proof systems that are efficient to be used in practice remains challenging. The most common and most efficient systems implemented are approaches based on sigma protocols, and approaches based on SNARKs (Succinct Non-interactive ARguments of Knowledge). Each approach has its own advantages and shortcomings, and are suited for certain statements.

  While sigma protocols are efficient for algebraic statements, they are expensive for non-algebraic statements. SNARKs, on the other hand, result in short proofs and efficient verification, and are better suited for proving statements about hash functions. But proving an algebraic statement, for instance, knowledge of discrete logarithm, is expensive as the prover needs to perform public-key operations proportional to the size of the circuit.

  Recent work achieve zero-knowledge proofs that are efficient for statements phrased as Boolean circuits based on Garbled circuits (GC). This, again, is expensive for large circuits, in addition to being inherently interactive. Thus, SNARKs and GC-based approaches are better suited for non-algebraic statements, and sigma protocols are efficient for algebraic statements.

  But in some applications, one is interested in proving combination statements, that is, statements that have both algebraic and non-algebraic components. The state of the art fails to take advantage of the best of all worlds and has to forgo the efficiency of one approach to obtain the other's. In this work, we ask how to efficiently prove a statement that is a combination of algebraic and non-algebraic statements.

  We first show how to combine the GC-based approach with sigma protocols. Then, we study how to combine sigma protocol proofs with SNARKs to obtain non-interactive arguments for combination statements. We show applications of our techniques to anonymous credentials, and privacy-preserving protocols on the blockchain. Finally, we study garbled circuits as a primitive and present an efficient way of hashing garbled circuits. We show applications of our hashing technique, including application to GC-based zero-knowledge.

• Ph.D. Thesis 2017 Circuit Complexity: New Techniques and Their Limitations Golovnev, Aleksandr Abstract | PDF

  Title: Circuit Complexity: New Techniques and Their Limitations

  Candidate: Golovnev, Aleksandr

  Advisor(s): Dodis, Yevgeniy; Regev, Oded

  Abstract:

  We study the problem of proving circuit lower bounds. The strongest known lower bound of 3n-o(n) for an explicit function was proven by Blum in 1984. We prove a lower bound of (3+1/86)n-o(n) for affine dispersers for sublinear dimensions.

  We introduce the weighted gate elimination method to give an elementary proof of a 3.11n lower bound for quadratic dispersers. (Although currently there are no explicit constructions of such functions.) Also, we develop a general framework which allows us to turn lower bounds proofs into upper bounds for Circuit SAT algorithms.

  Finally, we prove strong limitations of the developed techniques.

• Ph.D. Thesis 2017 Unsupervised Learning Under Uncertainty Mathieu, Michael Abstract | PDF

  Title: Unsupervised Learning Under Uncertainty

  Candidate: Mathieu, Michael

  Advisor(s): LeCun, Yann

  Abstract:

  Deep learning, in particular neural networks, achieved remarkable success in the recent years. However, most of it is based on supervised learning, and relies on ever larger datasets, and immense computing power. One step towards general artificial intelligence is to build a model of the world, with enough knowledge to acquire a kind of “common sense”. Representations learned by such a model could be reused in a number of other tasks. It would reduce the requirement for labeled samples and possibly acquire a deeper understanding of the problem. The vast quantities of knowledge required to build common sense preclude the use of supervised learning, and suggest to rely on unsupervised learning instead.

  The concept of uncertainty is central to unsupervised learning. The task is usually to learn a complex, multimodal distribution. Density estimation and generative models aim at representing the whole distribution of the data, while predictive learning consists of predicting the state of the world given the context and, more often than not, the prediction is not unique. That may be because the model lacks the capacity or the computing power to make a certain prediction, or because the future depends on parameters that are not part of the observation. Finally, the world can be chaotic of truly stochastic. Representing complex, multimodal continuous distributions with deep neural networks is still an open problem.

  In this thesis, we first assess the difficulties of representing probabilities in high dimensional spaces, and review the related work in this domain. We then introduce two methods to address the problem of video prediction, first using a novel form of linearizing auto-encoder and latent variables, and secondly using Generative Adversarial Networks (GANs). We show how GANs can be seen as trainable loss functions to represent uncertainty, then how they can be used to disentangle factors of variation. Finally, we explore a new non-probabilistic framework for GANs.

• Ph.D. Thesis 2017 Fine-scale Structure Design for 3D Printing Panetta, Francis Julian Abstract | PDF

  Title: Fine-scale Structure Design for 3D Printing

  Candidate: Panetta, Francis Julian

  Advisor(s): Zorin, Denis

  Abstract:

  Modern additive fabrication technologies can manufacture shapes whose geometric complexities far exceed what existing computational design tools can analyze or optimize. At the same time, falling costs have placed these fabrication technologies within the average consumer's reach. Especially for inexpert designers, new software tools are needed to take full advantage of 3D printing technology.

  My thesis develops such tools and demonstrates the exciting possibilities enabled by fine-tuning objects at the small scales achievable by 3D printing. The thesis applies two high-level ideas to invent these tools: two-scale design and worst-case analysis.

  The two-scale design approach addresses the problem that accurately simulating---let alone optimizing---geometry at the full resolution one can print requires orders of magnitude more computational power than currently available. However, we can use periodic homogenization to decompose the design problem into a small-scale problem (designing tileable structures achieving a particular deformation behavior) and a macro-scale problem (deciding where to place these structures in the larger object). We can then design structures for every possible deformation behavior and store them in a database, so that they can be re-used for many different macro-scale design problems.

  Worst-case analysis refers to determining how likely an object is to fracture by studying the worst possible scenario: the forces most efficiently breaking it. This analysis is needed when the designer has insufficient knowledge or experience to predict what forces an object will undergo, or when the design is intended for use in many different scenarios unknown a priori.

• Ph.D. Thesis 2017 On the Gaussian Measure Over Lattices Stephens-Davidowitz, Noah Abstract | PDF

  Title: On the Gaussian Measure Over Lattices

  Candidate: Stephens-Davidowitz, Noah

  Advisor(s): Dodis, Yevgeniy; Regev, Oded

  Abstract:

  We study the Gaussian mass of a lattice coset \[ \rho_s(\mathcal{L} - \vec{t}) := \sum_{\vec{y} \in \mathcal{L}} \exp(-\pi \|\vec{y} - \vec{t}\|^2/s^2) \; , \] where \(\mathcal{L} \subset \mathbb{R}^n\) is a lattice and \(\vec{t} \in \mathbb{R}^n\) is a vector describing a shift of the lattice. In particular, we use bounds on this Gaussian mass to obtain a partial converse to Minkowski's celebrated theorem bounding the number of lattice points in a ball.

  We also consider the discrete Gaussian distribution \(D_{\mathcal{L} - \vec{t}, s}\) induced by the Gaussian measure over \(\mathcal{L} - \vec{t}\), and we use procedures for sampling from this distribution to construct the current fastest known algorithms for the two most important computation problems over lattices, the Shortest Vector Problem (SVP) and the Closest Vector Problem (CVP).

  Finally, we study \(\rho_s(\mathcal{L} - \vec{t})\) and \(D_{\mathcal{L} - \vec{t}, s}\) as interesting computational and mathematical objects in their own right. In particular, we show that the computational problem of sampling from \(D_{\mathcal{L} - \vec{t}, s}\) is equivalent to CVP in a very strong sense (and that sampling from \(D_{\mathcal{L}, s}\) is no harder than SVP). We also prove a number of bounds on the moments of \(D_{\mathcal{L} - \vec{t}, s}\) and various monotonicity properties of \(\rho_s(\mathcal{L} - \vec{t})\).

• Ph.D. Thesis 2016 Decision Procedures for Finite Sets with Cardinality, and Local Theories Extensions Bansal, Kshitij Abstract | PDF

  Title: Decision Procedures for Finite Sets with Cardinality, and Local Theories Extensions

  Candidate: Bansal, Kshitij

  Advisor(s): Barrett, Clark; Wies, Thomas

  Abstract:

  Many tasks in design, verification, and testing of hardware and computer systems can be reduced to checking satisfiability of logical formulas. Certain fragments of first-order logic that model the semantics of prevalent data types, and hardware and software constructs, such as integers, bit-vectors, and arrays are thus of most interest. The appeal of satisfiability modulo theories (SMT) solvers is that they implement decision procedures for efficiently reasoning about formulas in these fragments. Thus, they can often be used off-the-shelf as automated back-end solvers in verification tools. In this thesis, we expand the scope of SMT solvers by developing decision procedures for new theories of interest in reasoning about hardware and software.

  First, we consider the theory of finite sets with cardinality. Sets are a common high-level data structure used in programming; thus, such a theory is useful for modeling program constructs directly. More importantly, sets are a basic construct of mathematics and thus natural to use when mathematically defining the properties of a computer system. We extend a calculus for finite sets to reason about cardinality constraints. The reasoning for cardinality involves tracking how different sets overlap. For an efficient procedure in an SMT solver, we'd like to avoid considering Venn regions directly, which has been the approach in earlier work. We develop a novel technique wherein potentially overlapping regions are considered incrementally. We use a graph to track the interaction of the different regions. Additionally, our technique leverages the procedure for reasoning about the other set operations (besides cardinality) in a modular fashion.

  Second, a limitation frequently encountered is that verification problems are often not fully expressible in the theories supported natively by the solvers. Many solvers allow the specification of application-specific theories as quantified axioms, but their handling is incomplete outside of narrow special cases. We show how SMT solvers can be used to obtain complete decision procedures for local theory extensions, an important class of theories that are decidable using finite instantiation of axioms. We present an algorithm that uses E-matching to generate instances incrementally during the search, significantly reducing the number of generated instances compared to eager instantiation strategies.

• Ph.D. Thesis 2016 Analyzing Source Code Across Static Conditionals Gazzillo, Paul Abstract | PDF

  Title: Analyzing Source Code Across Static Conditionals

  Candidate: Gazzillo, Paul

  Advisor(s): Wies, Thomas

  Abstract:

  We need better tools for C, such as source browsers, bug finders, and automated refactorings. The problem is that large C systems such as Linux are software product lines, containing thousands of configuration variables controlling every aspect of the software from architecture features to file systems and drivers. The challenge of such configurability is how do software tools accurately analyze all configurations of the source without the exponential explosion of trying them all separately. To this end, we focus on two key subproblems, parsing and the build system. The contributions of this thesis are the following: (1) a configuration-preserving preprocessor and parser called SuperC that preserves configurations in its output syntax tree; (2) a configuration-preserving Makefile evaluator called Kmax that collections Linux's compilation units and their configurations; and (3) a framework for configuration-aware analyses of source code using these tools.

  C tools need to process two languages: C itself and the preprocessor. The latter improves expressivity through file includes, macros, and static conditionals. But it operates only on tokens, making it hard to even parse both languages. SuperC is a complete, performant solution to parsing all of C. First, a configuration-preserving preprocessor resolves includes and macros yet leaves static conditionals intact, thus preserving a program's variability. To ensure completeness, we analyze all interactions between preprocessor features and identify techniques for correctly handling them. Second, a configuration-preserving parser generates a well-formed AST with static choice nodes for conditionals. It forks new subparsers when encountering static conditionals and merges them again after the conditionals. To ensure performance, we present a simple algorithm for table-driven Fork-Merge LR parsing and four novel optimizations. We demonstrate SuperC's effectiveness on the x86 Linux kernel.

  Large-scale C codebases like Linux are software product families, with complex build systems that tailor the software with myriad features. Such variability management is a challenge for tools, because they need awareness of variability to process all software product lines within the family. With over 14,000 features, processing all of Linux's product lines is infeasible by brute force, and current solutions employ incomplete heuristics. But having the complete set of compilation units with precise variability information is key to static tools such a bug-finders, which could miss critical bugs, and refactoring tools, since behavior-preservation requires a complete view of the software project. Kmax is a new tool for the Linux build system that extracts all compilation units with precise variability information. It processes build system files with a variability-aware \texttt{make} evaluator that stores variables in a conditional symbol table and hoists conditionals around complete statements, while tracking variability information as presence conditions. Kmax is evaluated empirically for correctness and completeness on the Linux kernel. Kmax is compared to previous work for correctness and running time, demonstrating that a complete solution's added complexity incurs only minor latency compared to the incomplete heuristic solutions.

  SuperC's configuration-preserving parsing of compilation units and Kmax's project-wide capabilities are in a unique position to process source code across all configurations. Bug-finding is one area where such capability is useful. Bugs may appear in untested combinations of configurations and testing each configuration one-at-a-time is infeasible. For example, one compilation units that defines a global function called by other compilation units may not be linked into the final program due to configuration variable selection. Such a bug can be found with Kmax and SuperC's cross-configuration capability. Cilantro is a framework for creating variability-aware bug-checkers. Kmax is used to determine the complete set of compilation units and the combinations of features that activate them, while SuperC's parsing framework is extended with semantic actions in order implement the checkers. A checker for linker errors across all compilation units in the Linux kernel demonstrates each part of the Cilantro framework and is evaluated on the Linux source code.

• Ph.D. Thesis 2016 Semi-Supervised Learning for Electronic Phenotyping in Support of Precision Medicine Halpern, Yonatan Abstract | PDF

  Title: Semi-Supervised Learning for Electronic Phenotyping in Support of Precision Medicine

  Candidate: Halpern, Yonatan

  Advisor(s): Sontag, David

  Abstract:

  Medical informatics plays an important role in precision medicine, delivering the right information to the right person, at the right time. With the introduction and widespread adoption of electronic medical records, in the United States and world-wide, there is now a tremendous amount of health data available for analysis. Electronic record phenotyping refers to the task of determining, from an electronic medical record entry, a concise descriptor of the patient, comprising of their medical history, current problems, presentation, etc. In inferring such a phenotype descriptor from the record, a computer, in a sense, "understands", the relevant parts of the record. These phenotypes can then be used in downstream applications such as cohort selection for retrospective studies, real-time clinical decision support, contextual displays, intelligent search, and precise alerting mechanisms.

  To handle the incomplete data present in medical records, we use the formal framework of probabilistic graphical models with latent or unobserved variables. The first part of the thesis presents two different structural conditions under which learning with latent variables is computationally tractable. The first is the "anchored" condition, where every latent variable has at least one child that is not shared by any other parent. The second is the "singly-coupled" condition, where every latent variable is connected to at least three children that satisfy conditional independence (possibly after a transformation of the data). Variables that satisfy these conditions can be specified by an expert without requiring that the entire structure or its parameters be specified, allowing for effective use of human expertise and making room for statistical learning to do some of the heavy lifting in model learning. For both the anchored and singly-coupled conditions, practical algorithms are presented.

  The second part of the thesis describes real-life applications using the anchored condition for electronic phenotyping. A human-in-the-loop learning system and a functioning emergency informatics system for real-time extraction of important clinical variables are described and evaluated.

  The algorithms and discussion presented here were developed for the purpose of improving healthcare, but are much more widely applicable, dealing with the very basic questions of identifiability and learning models with latent variables - a problem that lies at the very heart of the natural and social sciences.

• Ph.D. Thesis 2016 Improving Knowledge Base Population with Information Extraction Li, Xiang Abstract | PDF

  Title: Improving Knowledge Base Population with Information Extraction

  Candidate: Li, Xiang

  Advisor(s): Grishman, Ralph

  Abstract:

  Knowledge Bases (KBs) are data resources that encode world knowledge in machine-readable formats. Knowledge Base Population (KBP) aims at understanding this knowledge and extending KBs with more semantic information, which is a fundamental problem in Artificial Intelligence. It can benefit a wide range of tasks, such as semantic search and question answering. Information Extraction (IE), the task of discovering important types of facts (entities, relations and events) in unstructured text, is necessary and crucial for successfully populating knowledge bases. This dissertation focuses on four essential aspects of knowledge base population by leveraging IE techniques: extracting facts from unstructured data, validating the extracted information, accelerating and enhancing systems with less annotation effort, and utilizing knowledge bases to improve real-world applications.

  First, we investigate the Slot Filling task, which is a key component for knowledge base population. Slot filling aims to collect information from a large collection of news, web, or other sources of documents to determine a set of predefined attributes ("slots") for given person and organization entities. We introduce a statistical language understanding approach to automatically construct personal (user-centric) knowledge bases from conversational dialogs.

  Second, we consider how to probabilistically estimate the correctness of the extracted slot values. Despite the significant progress of KBP research and systems in recent years, slot filling approaches are still far from completely reliable. Using the NIST KBP Slot Filling task as a case study, we propose a confidence estimation model based on the Maximum Entropy framework, and demonstrate the effectiveness of this model in both precision and the capability to improve the slot filling aggregation through a weighted voting strategy.

  Third, we study rich annotation guided learning to fill the gap between an expert annotator and a feature engineer. We develop an algorithm to enrich features with the guidance of all levels of rich annotations from human annotators. We also evaluate the comparative efficacy, generality and scalability of this framework by conducting case studies on three distinct applications in various domains, including facilitating KBP slot filling systems. Empirical studies demonstrate that with little additional annotation time, we can significantly improve the performance for all tasks.

  Finally, we explore utilizing knowledge bases in a real-world application - personalized content recommendation. Traditional systems infer user interests from surface-level features derived from online activity logs and user demographic profiles, rather than deeply understanding the context semantics. We conduct a systematic study to show the effectiveness of incorporating deep semantic knowledge encoded in the entities on modeling user interests, by utilizing the abundance of entity information from knowledge bases.

• Ph.D. Thesis 2016 Improving SAT Solvers by Exploiting Empirical Characteristics of CDCL Oh, Chanseok Abstract | PDF

  Title: Improving SAT Solvers by Exploiting Empirical Characteristics of CDCL

  Candidate: Oh, Chanseok

  Advisor(s): Wies, Thomas

  Abstract:

  The Boolean Satisfiability Problem (SAT) is a canonical decision problem originally shown to be NP-complete in Cook’s seminal work on the theory of computational complexity. The SAT problem is one of several computational tasks identified by researchers as core problems in computer science. The existence of an efficient decision procedure for SAT would imply P = NP. However, numerous algorithms and techniques for solving the SAT problem have been proposed in various forms in practical settings. Highly efficient solvers are now actively being used, either directly or as a core engine of a larger system, to solve real-world problems that arise from many application domains. These state-of-the-art solvers use the Davis-Putnam-Logemann-Loveland (DPLL) algorithm extended with ConflictDriven Clause Learning (CDCL). Due to the practical importance of SAT, building a fast SAT solver can have a huge impact on current and prospective applications. The ultimate contribution of this thesis is improving the state of the art of CDCL by understanding and exploiting the empirical characteristics of how CDCL works on real-world problems. The first part of the thesis shows empirically that most of the unsatisfiable real-world problems solvable by CDCL have a refutation proof with near-constant width for the great portion of the proof. Based on this observation, the thesis provides an unconventional perspective that CDCL solvers can solve real-world problems very efficiently and often more efficiently just by maintaining a small set of certain classes of learned clauses. The next part of the thesis focuses on understanding the inherently different natures of satisfiable and unsatisfiable problems and their implications on the empirical workings of CDCL. We examine the varying degree of roles and effects of crucial elements of CDCL based on the satisfiability status of a problem. Ultimately, we propose effective techniques to exploit the new insights about the different natures of proving satisfi- ability and unsatisfiability to improve the state of the art of CDCL. In the last part of the thesis, we present a reference solver that incorporates all the techniques described in the thesis. The design of the presented solver emphasizes minimality in implementation while guaranteeing state-of-the-art performance. Several versions of the reference solver have demonstrated top-notch performance, earning several medals in the annual SAT competitive events. The minimal spirit of the reference solver shows that a simple CDCL framework alone can still be made competitive with state-of-the-art solvers that implement sophisticated techniques outside the CDCL framework.

• Ph.D. Thesis 2016 Graph-based Approaches to Resolve Entity Ambiguity Pershina, Maria Abstract | PDF

  Title: Graph-based Approaches to Resolve Entity Ambiguity

  Candidate: Pershina, Maria

  Advisor(s): Grishman, Ralph

  Abstract:

  Information Extraction is the task of automatically extracting structured information from unstructured or semi-structured machine-readable documents. One of the challenges of Information Extraction is to resolve ambiguity between entities either in a knowledge base or in text documents. There are many variations of this problem and it is known under different names, such as coreference resolution, entity disambiguation, entity linking, entity matching, etc. For example, the task of coreference resolution decides whether two expressions refer to the same entity; entity disambiguation determines how to map an entity mention to an appropriate entity in a knowledge base (KB); the main focus of entity linking is to infer that two entity mentions in a document(s) refer to the same real world entity even if they do not appear in a KB; entity matching (also record deduplication, entity resolution, reference reconciliation) is to merge records from databases if they refer to the same object.

  Resolving ambiguity and finding proper matches between entities is an important step for many downstream applications, such as data integration, question answering, relation extraction, etc. The Internet has enabled the creation of a growing number of large-scale knowledge bases in a variety of domains, posing a scalability challenge for Information Extraction systems. Tools for automatically aligning these knowledge bases would make it possible to unify many sources of structured knowledge and to answer complex queries. However the efficient alignment of large-scale knowledge bases still poses a considerable challenge.

  Various aspects and different settings to resolve ambiguity between entities are studied in this dissertation. A new scalable domain-independent graph-based approach utilizing Personalized Page Rank is developed for entity matching across large-scale knowledge bases and evaluated on datasets of 110 million and 203 million entities. A new model for entity disambiguation between a document and a knowledge base utilizing a document graph and effectively filtering out noise is proposed. A new technique based on a paraphrase detection model is proposed to recognize name variations for an entity in a document. A new approach integrating a graph-based entity disambiguation model and this technique is presented for an entity linking task and is evaluated on a dataset for  the Text Analysis Conference Entity Discovery and Linking 2014 task.

• Ph.D. Thesis 2016 Partition Memory Models for Program Analysis Wang, Wei Abstract | PDF

  Title: Partition Memory Models for Program Analysis

  Candidate: Wang, Wei

  Advisor(s): Barrett, Clark; Wies, Thomas

  Abstract:

  Scalability is a key challenge in static program analyses based on solvers for Satisfiability Modulo Theories (SMT). For imperative languages like C, the approach taken for modeling memory can play a significant role in scalability. The main theme of this thesis is using partitioned memory models to divide up memory based on the alias information derived from a points-to analysis.

  First, a general analysis framework based on memory partitioning is presented. It incorporates a points-to analysis as a preprocessing step to determine a conservative approximation of which areas of memory may alias or overlap and splits the memory into distinct arrays for each of these areas.

  Then we propose a new cell-based field-sensitive points-to analysis, which is an extension of Steensgaard’s unification-based algorithms. A cell is a unit of access with scalar or record type. Arrays and dynamically memory allocations are viewed as a collection of cells. We show how our points-to analysis yields more precise alias information for programs with complex heap data structures.

  Our work is implemented in Cascade, a static analysis framework for C programs. It replaces the former flat memory model that models the memory as a single array of bytes. We show that the partitioned memory models achieve better scalability within Cascade, and the cell-based memory model, in particular, improves the performance significantly, making Cascade a state-of-the-art C analyzer.

• Ph.D. Thesis 2016 Learning Algorithms from Data Zaremba, Wojciech Abstract | PDF

  Title: Learning Algorithms from Data

  Candidate: Zaremba, Wojciech

  Advisor(s): Fergus, Rob; LeCun, Yann

  Abstract:

  Statistical machine learning is concerned with learning models that describe observations. We train our models from data on tasks like machine translation or object recognition because we cannot explicitly write down programs to solve such problems. A statistical model is only useful when it generalizes to unseen data. Solomonoff has proved that one should choose the model that agrees with the observed data, while preferring the model that can be compressed the most, because such a choice guarantees the best possible generalization. The size of the best possible compression of the model is called the Kolmogorov complexity of the model. We define an algorithm as a function with small Kolmogorov complexity.

  This Ph.D. thesis outlines the problem of learning algorithms from data and shows several partial solutions to it. Our data model is mainly neural networks as they have proven to be successful in various domains like object recognition, language modeling, speech recognition and others. First, we examine empirical trainability limits for classical neural networks. Then, we extend them by providing interfaces, which provide a way to read memory, access the input, and postpone predictions. The model learns how to use them with reinforcement learning techniques like Reinforce and Q-learning. Next, we examine whether contemporary algorithms such as convolution layer can be automatically rediscovered. We show that it is possible indeed to learn convolution as a special case in a broader range of models. Finally, we investigate whether it is directly possible to enumerate short programs and find a solution to a given problem. This follows the original line of thought behind the Solomonoff induction. Our approach is to learn a prior over programs such that we can explore them efficiently.

• Ph.D. Thesis 2016 Distributed Stochastic Optimization for Deep Learning Zhang, Sixin Abstract | PDF

  Title: Distributed Stochastic Optimization for Deep Learning

  Candidate: Zhang, Sixin

  Advisor(s): LeCun, Yann

  Abstract:

  We study the problem of how to distribute the training of large-scale deep learning models in the parallel computing environment. We propose a new distributed stochastic optimization method called Elastic Averaging SGD (EASGD). We analyze the convergence rate of the EASGD method in the synchronous scenario and compare its stability condition with the existing ADMM method in the round-robin scheme. An asynchronous and momentum variant of the EASGD method is applied to train deep convolutional neural networks for image classification on the CIFAR and ImageNet datasets. Our approach accelerates the training and furthermore achieves better test accuracy. It also requires a much smaller amount of communication than other common baseline approaches such as the DOWNPOUR method.

  We then investigate the limit in speedup of the initial and the asymptotic phase of the mini-batch SGD, the momentum SGD, and the EASGD methods. We find that the spread of the input data distribution has a big impact on their initial convergence rate and stability region. We also find a surprising connection between the momentum SGD and the EASGD method with a negative moving average rate. A non-convex case is also studied to understand when EASGD can get trapped by a saddle point.

  Finally, we scale up the EASGD method by using a tree structured network topology. We show empirically its advantage and challenge. We also establish a connection between the EASGD and the DOWNPOUR method with the classical Jacobi and the Gauss-Seidel method, thus unifying a class of distributed stochastic optimization methods.

• Ph.D. Thesis 2016 Pushing the Limits of Additive Fabrication Technologies Zhou, Qingnan (James) Abstract | PDF

  Title: Pushing the Limits of Additive Fabrication Technologies

  Candidate: Zhou, Qingnan (James)

  Advisor(s): Zorin, Denis

  Abstract:

  A rough symmetry can be observed in the stock price of 3D Systems (NYSE:DDD), the leading and largest 3D printer manufacturer, from its IPO on June 3, 2011 to the beginning of 2016. The price sky rocketed nearly 600% from 2011 to the end of 2013, and took a free fall back to its original value by 2016. Coincidentally, it is also the period during which I got my hands dirty and investigated some of the toughest challenges as well as exciting new possibilities associated with different types of 3D printing technologies. In this thesis, I documented my attempts from 3 different angles to push the limits of 3D printing: printability, microstructure design and robust geometry processing with mesh arrangements.

  Printability check has long been the bottleneck that prevents 3D printing from scaling up. Oftentimes, designers of 3D models lack the expertise or tools to ensure 3D printability. 3D printing service providers typically rely human inspections to filter out unprintable designs. This process is manual and error-prone. As designs become ever more complex, manual printability check becomes increasingly difficult. To tackle this problem, my colleagues and I proposed an algorithm to automatically determine structurally weak regions and the worst-case usage scenario to break a given model. We validate the algorithm by physically break a number of real 3D printed designs.

  A key distinctive feature of 3D printing technologies is that the cost and time of fabrication is uncorrelated with geometric complexity. This opens up many exciting new possibilities. In particular, by pushing geometric complexity to the extreme, 3D printing has the potential of fabricating soft, deformable shapes with microscopic structures using a single raw material. In our recent SIGGRAPH publication, my colleagues and I have not only demonstrated fabricating microscopic frame structures is possible but also proposed an entire pipeline for designing spatially varying microstructures to satisfy target material properties or deformation goals.

  With the boost of 3D printing technologies, 3D models have become more abundant and easily accessible than ever before. These models are sometimes known as "wild" models because they differ significantly in complexity and quality from traditional models in graphics researches. This poses a serious challenge in robustly analyzing 3D designs. Many state-of-the-art geometry processing algorithms/libraries are ill-prepared for dealing with "wild" models that are non-manifold, self-intersecting, locally degenerate and/or containing multiple and possibly nested components. In our most recent SIGGRAPH submission, we proposed a systematic recipe based on mesh arrangements for conducting a family of exact constructive solid geometry operations. We exhaustively tested our algorithm on 10,000 "wild" models crawled from Thingiverse, a popular online shape repository. Both the code and the dataset are freely available to the public.

• Ph.D. Thesis 2015 Big Data Analytics for Development: Events, Knowledge Graphs and Predictive Models Chakraborty, Sunandan Abstract | PDF

  Title: Big Data Analytics for Development: Events, Knowledge Graphs and Predictive Models

  Candidate: Chakraborty, Sunandan

  Advisor(s): Subramanian, Lakshminarayanan; Nyarko, Yaw

  Abstract:

  Volatility in critical socio-economic indices can have a significant negative impact on global development. This thesis presents a suite of novel big data analytics algorithms that operate on unstructured Web data streams to automatically infer events, knowledge graphs and predictive models to understand, characterize and predict the volatility of socioeconomic indices.

  This thesis makes four important research contributions. First, given a large volume of diverse unstructured news streams, we present new models for capturing events and learning spatio-temporal characteristics of events from news streams. We specifically explore two types of event models in this thesis: one centered around the concept of event triggers and a probabilistic meta-event model that explicitly delineates named entities from text streams to learn a generic class of meta-events. The second contribution focuses on learning several different types of knowledge graphs from news streams and events: a) Spatio-temporal article graphs capture intrinsic relationships between different news articles; b) Event graphs characterize relationships between events and given a news query, provide a succinct summary of a timeline of events relating to a query; c) Event-phenomenon graphs that provide a condensed representation of classes of events that relate to a given phenomena at a given location and time; d) Causality testing on word-word graphs which can capture strong spatio-temporal relationships between word occurrences in news streams; e) Concept graphs that capture relationships between different word concepts that occur in a given text stream.

  The third contribution focuses on connecting the different knowledge graph representations and structured time series data corresponding to a socio-economic index to automatically learn event-driven predictive models for the given socio-economic index to predict future volatility. We propose several types of predictive models centered around our two event models: event triggers and probabilistic meta-events. The final contribution focuses on a broad spectrum of inference case studies for different types of socio-economic indices including food prices, stock prices, disease outbreaks and interest rates. Across all these indices, we show that event-driven predictive models provide significant improvements in prediction accuracy over state-of-the-art techniques.

• Ph.D. Thesis 2015 SMT-Based and Disjunctive Relational Abstract Domains for StaticAnalysis Chen, Junjie Abstract | PDF

  Title: SMT-Based and Disjunctive Relational Abstract Domains for StaticAnalysis

  Candidate: Chen, Junjie

  Advisor(s): Patrick Cousot

  Abstract:

  Abstract Interpretation is a theory of sound approximation of program semantics. In recent decades, it has been widely and successfully applied to the static analysis of computer programs. In this thesis, we will work on abstract domains, one of the key concepts in abstract interpretation, which aim at automatically collecting information about the set of all possible values of the program variables. We will focus, in particularly, on two aspects: the combination with theorem provers and the refinement of existing abstract domains.

  Satisfiability modulo theories (SMT) solvers are popular theorem provers, which proved to be very powerful tools for checking the satisfiability of first-order logical formulas with respect to some background theories. In the first part of this thesis, we introduce two abstract domains whose elements are logical formulas involving finite conjunctions of affine equalities and finite conjunctions of linear inequalities. These two abstract domains rely on SMT solvers for the computation of transformations and other logical operations.

  In the second part of this thesis, we present an abstract domain functor whose elements are binary decision trees. It is parameterized by decision nodes which are a set of boolean tests appearing in the programs and by a numerical or symbolic abstract domain whose elements are the leaves. This new binary decision tree abstract domain functor provides a flexible way of adjusting the cost/precision ratio in path-dependent static analysis.

• Ph.D. Thesis 2015 Iris: Mitigating Phase Noise in Millimeter Wave OFDM Systems Dhananjay, Aditya Abstract | PDF

  Title: Iris: Mitigating Phase Noise in Millimeter Wave OFDM Systems

  Candidate: Dhananjay, Aditya

  Advisor(s): Li, Jinyang

  Abstract:

  Next-generation wireless networks are widely expected to operate over millimeter-wave (mmW) frequencies of over 28GHz. These bands mitigate the acute spectrum shortage in the conventional microwave bands of less than 6GHz. The shorter wavelengths in these bands also allow for building dense antenna arrays on a single chip, thereby enabling various MIMO configurations and highly directional links that can increase the spatial reuse of spectrum.

  While attempting to build a practical over-the-air (OTA) link over mmW, we realized that the traditional baseband processing techniques used in the microwave bands simply could not cope with the exacerbated frequency offsets (or phase noise) observed in the RF oscillators at these bands. While the frequency offsets are large, the real difficulty arose from the fact that they varied significantly over very short time-scales.Traditional feedback loop techniques still left significant residual offsets, which in turn led to inter-carrier-interference (ICI). The result was very high symbol error rates (SER).

  This thesis presents Iris, a baseband processing block that enables clean mmW links, even in the presence of previously fatal amounts of phase noise. Over real mmW hardware, Iris reduces the SER by one to two orders of magnitude, as compared to competing techniques.

• Ph.D. Thesis 2015 Predicting Images using Convolutional Networks: Visual Scene Understanding with Pixel Maps Eigen, David Abstract | PDF

  Title: Predicting Images using Convolutional Networks: Visual Scene Understanding with Pixel Maps

  Candidate: Eigen, David

  Advisor(s): Fergus, Rob

  Abstract:

  In the greater part of this thesis, we develop a set of convolutional networks that infer predictions at each pixel of an input image. This is a common problem that arises in many computer vision applications: For example, predicting a semantic label at each pixel describes not only the image content, but also fine-grained locations and segmenta- tions; at the same time, finding depth or surface normals provide 3D geometric relations between points. The second part of this thesis investigates convolutional models also in the contexts of classification and unsupervised learning.

  To address our main objective, we develop a versatile Multi-Scale Convolutional Network that can be applied to diverse vision problems using simple adaptations, and apply it to predict depth at each pixel, surface normals and semantic labels. Our model uses a series of convolutional network stacks applied at progressively finer scales. The first uses the entire image field of view to predict a spatially coarse set of feature maps based on global relations; subsequent scales correct and refine the output, yielding a high resolution prediction. We look exclusively at depth prediction first, then generalize our method to multiple tasks. Our system achieves state-of-the-art results on all tasks we investigate, and can match many image details without the need for superpixelation.

  Leading to our multi-scale network, we also design a purely local convolutional network to remove dirt and raindrops present on a window surface, which learns to identify and inpaint compact corruptions. We also we investigate a weighted nearest-neighbors labeling system applied to superpixels, in which we learn weights for each example, and use local context to find rare class instances.

  In addition, we investigate the relative importance of sizing parameters using a recursive convolutional network, finding that network depth is most critical. We also develop a Convolutional LISTA Autoencoder, which learns features similar to stacked sparse coding at a fraction of the cost, combine it with a local entropy objective, and describe a convolutional adaptation of ZCA whitening.

• Ph.D. Thesis 2015 Unsupervised Feature Learning in Computer Vision Goroshin, Ross Abstract | PDF

  Title: Unsupervised Feature Learning in Computer Vision

  Candidate: Goroshin, Ross

  Advisor(s): LeCun, Yann

  Abstract:

  Much of computer vision has been devoted to the question of representation through feature extraction. Ideal features transform raw pixel intensity values to a representation in which common problems such as object identification, tracking, and segmentation are easier to solve. Recently, deep feature hierarchies have proven to be immensely successful at solving many problems in computer vision. In the supervised setting, these hierarchies are trained to solve specific problems by minimizing an objective function of the data and problem specific label information. Recent findings suggest that despite being trained on a specific task, the learned features can be transferred across multiple visual tasks. These findings suggests that there exists a generically useful feature representation for natural visual data.

  This work aims to uncover the principles that lead to these generic feature representations in the unsupervised setting, which does not require problem specific label information. We begin by reviewing relevant prior work, particularly the literature on autoencoder networks and energy based learning. We introduce a new regularizer for autoencoders that plays an analogous role to the partition function in probabilistic graphical models. Next we explore the role of specialized encoder architectures for sparse inference. The remainder of the thesis explores visual feature learning from video. We establish a connection between slow-feature learning and metric learning, and experimentally demonstrate that semantically coherent metrics can be learned from natural videos. Finally, we posit that useful features linearize natural image transformations in video. To this end, we introduce a new architecture and loss for training deep feature hierarchies that linearize the transformations observed in unlabeled natural video sequences by learning to predict future frames in the presence of uncertainty.

• Ph.D. Thesis 2015 Efficient and Trustworthy Theory Solver for Bit-vectors in SatisfiabilityModulo Theories Hadarean, Liana Abstract | PDF

  Title: Efficient and Trustworthy Theory Solver for Bit-vectors in SatisfiabilityModulo Theories

  Candidate: Hadarean, Liana

  Advisor(s): Barrett, Clark

  Abstract:

  As software and hardware systems grow in complexity, automated techniques for ensuring their correctness are becoming increasingly important. Many modern formal verification tools rely on back-end satisfiability modulo theories (SMT) solvers to discharge complex verification goals. These goals are usually formalized in one or more fixed first-order logic theories, such as the theory of fixed-width bit-vectors. The theory of bit-vectors offers a natural way of encoding the precise semantics of typical machine operations on binary data. The predominant approach to deciding the bit-vector theory is via eager reduction to propositional logic. While this often works well in practice, it does not scale well as the bit-width and number of operations increase. The first part of this thesis seeks to fill this gap, by exploring efficient techniques of solving bit-vector constraints that leverage the word-level structure. We propose two complementary approaches: an eager approach that takes full advantage of the solving power of off the shelf propositional logic solvers, and a lazy approach that combines on-the-fly algebraic reasoning with efficient propositional logic solvers. In the second part of the thesis, we propose a proof system for encoding automatically checkable refutation proofs in the theory of bit-vectors. These proofs can be automatically generated by the SMT solver, and act as a certificate for the correctness of the result.

• Ph.D. Thesis 2015 Predicting the Market Value of Single-Family Residences Lowrance, Roy Abstract | PDF

  Title: Predicting the Market Value of Single-Family Residences

  Candidate: Lowrance, Roy

  Advisor(s): LeCun, Yann; Shasha, Dennis

  Abstract:

  This work develops the best linear model of residential real estate prices for 2003 through 2009 in Los Angeles County. It differs from other studies comparing models for predicting house prices by covering a larger geographic area than most, more houses than most, a longer time period than most, and the time period both before and after the real estate price boom in the United States. 

  In addition, it open sources all of the software. We test designs for linear models to determine the best form for the model as well as the training period, features, and regularizer that produce the lowest errors. We compare the best of our linear models to random forests and point to directions for further research.

• Ph.D. Thesis 2015 Building Fast, CPU-Efficient Distributed Systems on Ultra-Low Latency, RDMA-Capable Networks Mitchell, Christopher Abstract | PDF

  Title: Building Fast, CPU-Efficient Distributed Systems on Ultra-Low Latency, RDMA-Capable Networks

  Candidate: Mitchell, Christopher

  Advisor(s): Li, Jinyang

  Abstract:

  Modern datacenters utilize traditional Ethernet interconnects to connect hundreds or thousands of machines. Although inexpensive and ubiquitous, Ethernet imposes design constraints on datacenter-scale distributed storage systems that use traditional client-server architectures. Recent technological trends indicate that future datacenters will embrace interconnects with ultra-low latency, high bandwidth, and the ability to offload work from servers to clients. Future datacenter-scale distributed storage systems will need to be designed specifically to exploit these features. This thesis explores what these features mean for large-scale in-memory storage systems, and derives two key insights for building RDMA-aware distributed systems.

  First, relaxing locality between data and computation is now practical: data can be copied from servers to clients for computation. Second, selectively relaxing data-computation locality makes it possible to optimally balance load between server and client CPUs to maintain low application latency. This thesis presents two in-memory distributed storage systems built around these two insights, Pilaf and Cell, that demonstrate effective use of ultra-low-latency, RDMA-capable interconnects. Through Pilaf and Cell, this thesis demonstrates that by combining RDMA and message passing to selectively relax locality, systems can achieve ultra-low latency and optimal load balancing with modest CPU resources.

• Ph.D. Thesis 2015 Instance Segmentation of RGBD Scenes Silberman, Nathan Abstract | PDF

  Title: Instance Segmentation of RGBD Scenes

  Candidate: Silberman, Nathan

  Advisor(s): Fergus, Rob

  Abstract:

  The vast majority of literature in scene parsing can be described as semantic pixel labeling or semantic segmentation: predicting the semantic class of the object represented by each pixel in the scene. Our familiar perception of the world, however, provides a far richer representation. Firstly, rather than just being able to predict the semantic class of a location in a scene, humans are able to reason about object instances. Discriminating between a region that might represent a single object versus ten objects is a crucial and basic faculty. Secondly, rather than reasoning about objects as merely occupying the space visible from a single vantage point, we are able to quickly and easily reason about an object's true extent in 3D. Thirdly, rather than viewing a scene as a collection of objects independently existing in space, humans exhibit a representation of scenes that is highly grounded through a intuitive model of physics. Such models allow us to reason about how objects relate physically: via physical support relationships.

  Instance segmentation is the task of segmenting a scene into regions which correspond to individual object instances. We argue that this task is not only closer to our own perception of the world than semantic segmentation, but also directly allows for subsequent reasoning about a scenes constituent elements. We explore various strategies for instance segmentation in indoor RGBD scenes.

  Firstly, we explore tree-based instance segmentation algorithms. The utility of trees for semantic segmentation has been thoroughly demonstrated and we adapt them to instance segmentation and analyze both greedy and global approaches to inference.

  Next, we investigate exemplar-based instance segmentation algorithms, in which a set of representative exemplars are chosen from a large pool of regions and pixels are assigned to exemplars. Inference can either be performed in two stages, exemplar selection followed by pixel-to-exemplar assignment, or in a single joint reasoning stage. We consider the advantages and disadvantages of each approach.

  We introduce the task of support-relation prediction in which we predict which objects are physically supporting other objects. We propose an algorithm and a new set of features for performing discriminative support prediction, we demonstrate the effectiveness of our method and compare training mechanisms.

  Finally, we introduce an algorithm for inferring scene and object extent. We demonstrate how reasoning about 3D extent can be done by extending known 2D methods and highlight the strengths and limitations of this approach.

• Ph.D. Thesis 2015 Localization of Humans in Images Using Convolutional Networks Tompson, Jonathan Abstract | PDF

  Title: Localization of Humans in Images Using Convolutional Networks

  Candidate: Tompson, Jonathan

  Advisor(s): Bregler, Christopher

  Abstract:

  Tracking of humans in images is a long standing problem in computer vision research for which, despite significant research effort, an adequate solution has not yet emerged. This is largely due to the fact that human body localization is complicated and difficult; potential solutions must find the location of body joints in images with invariance to shape, lighting and texture variation and it must do so in the presence of occlusion and incomplete data. However, despite these significant challenges, this work will present a framework for human body pose localization that not only offers a significant improvement over existing traditional architectures, but has sufficient localization performance and computational efficiency for use in real-world applications.

  At it's core, this framework makes use of Convolutional Networks to infer the location of body joints efficiently and accurately. We describe solutions to two applications 1) hand-tracking from a depth image source and 2) human body-tracking from and RGB image source. For both these applications we show that Convolutional Networks are able to significantly out-perform existing state-of-the-art.

  We propose a new hybrid architecture that consists of a deep Convolutional Network and a Probabilistic Graphical Model which can exploit structural domain constraints such as geometric relationships between body joint locations to improve tracking performance. We then explore the use of both color and motion features to improve tracking performance. Finally we introduce a novel architecture which includes an efficient ‘position refinement’ model that is trained to estimate the joint offset location within a small region of the image. This refinement model allows our network to improve spatial localization accuracy even with large amounts of spatial pooling.

• Ph.D. Thesis 2015 Joint Training of a Neural Network and a Structured Model for Computer Vision Wan, Li Abstract | PDF

  Title: Joint Training of a Neural Network and a Structured Model for Computer Vision

  Candidate: Wan, Li

  Advisor(s): Fergus, Rob

  Abstract:

  Identifying objects and telling where they are in real world images is one of the most important problems in Artificial Intelligence. The problem is challenging due to: occluded objects, varying object viewpoints and object deformations. This makes the vision problem extremely difficult and cannot be efficiently solved without learning.

  This thesis explores hybrid systems that combine a neural network as a trainable feature extractor and structured models that capture high level information such as object parts. The resulting models combine the strengths of the two approaches: a deep neural network which provides a powerful non-linear feature transformation and a high level structured model which integrates domain-specific knowledge. We develop discriminative training algorithms to jointly optimize these entire models end-to-end.

  First, we proposed a unified model which combines a deep neural network with a latent topic model for image classification. The hybrid model is shown to outperform models based solely on neural networks or topic model alone. Next, we investigate techniques for training a neural network system, introducing an effective way of regularizing the network called DropConnect. DropConnect allows us to train large models while avoiding over-fitting. This yields state-of-the-art results on a variety of standard benchmarks for image classification. Third, we worked on object detection for PASCAL challenge. We improved the deformable parts model and proposed a new non-maximal suppression algorithm. This system was the joint winner of the 2011 challenge. Finally, we develop a new hybrid model which integrates a deep network, deformable parts model and non-maximal suppression. Joint training of our hybrid model shows clear advantage over train each component individually, and achieving competitive result on standard benchmarks.

• Ph.D. Thesis 2015 Partition Memory Models in Program Analysis Wang, Wei Abstract | PDF

  Title: Partition Memory Models in Program Analysis

  Candidate: Wang, Wei

  Advisor(s): Barrett, Clark

  Abstract:

  Scalability is a key challenge in static program analyses based on solvers for Satisfiability Modulo Theories (SMT). For imperative languages like C, the approach taken for modeling memory can play a significant role in scalability. The main theme of this thesis is using partitioned memory models to divide up memory based on the alias information derived from a points-to analysis.

  First, a general analysis framework based on memory partitioning is presented. It incorporates a points-to analysis as a preprocessing step to determine a conservative approximation of which areas of memory may alias or overlap and splits the memory into distinct arrays for each of these areas.

  Then we propose a new cell-based field-sensitive points-to analysis, which is an extension of Steensgaard's unification-based algorithms. A cell is a unit of access with scalar or record type. Arrays and dynamically memory allocations are viewed as a collection of cells. We show how our points-to analysis yields more precise alias information for programs with complex heap data structures.

  Our work is implemented in Cascade, a static analysis framework for C programs. It replaces the former at memory model that models the memory as a single array of bytes. We show that the partitioned memory models achieve better scalability within Cascade, and the cell-based memory model, in particular, improves the performance significantly, making Cascade a state-of-the-art C analyzer.

• Ph.D. Thesis 2014 On the Human Form: Efficient acquisition, modeling and manipulation of thehuman body Braga, Otavio Abstract | PDF

  Title: On the Human Form: Efficient acquisition, modeling and manipulation of thehuman body

  Candidate: Braga, Otavio

  Advisor(s): Geiger, Davi

  Abstract:

  This thesis concerns the acquisition, modeling and manipulation of the human form.

  First, we acquire body models. We introduce an efficient bootstraped algorithm that we employed to register over 2,000 high resolution body scans of male and female adult subjects. Our algorithm outputs not only the traditional vertex correspondences, but also directly produces a high quality model which can be immediately deformed. We then employ the result to fit noisy depth maps coming from now commercially available 3D sensors such as Microsoft's Kinect and PrimeSense's Carmine.

  We conclude by describing a new real-time system for image-based body manipulation called BodyJam, that lets you change your outfit with a finger snap. BodyJam is inspired by a technique invented by the surrealists a century ago: "Exquisite corpse", a method by which a collection of images (of body parts) is collectively assembled. BodyJam does it on a video display that mirrors the pose in real-time of a real-person standing in front of the camera/display mirror, and allows the user to change clothes and other appearance attributes. Using Microsoft's Kinect, poses are matched to a video database of different torsos and legs, and "pages" showing different clothes are turned by handwitch focus to the topic of body manipulation. We first revisit the more traditional way of specifying bodies from a set of measurements, such as coming from clothing sizing charts, showing how the statistics of the population learned during the registration can aid us in accurately defining the body shape. We then introduce a new manipulation metaphor, where we navigate through the space of body shapes and poses by directly dragging the body mesh surface.

  We conclude by describing a new real-time system for image-based body manipulation called BodyJam, that lets you change your outfit with a finger snap. BodyJam is inspired by a technique invented by the surrealists a century ago: "Exquisite Corpse", a method by which a collection of images (of body parts) is collectively assembled. BodyJam does it on a video display that mirrors the pose in real-time of a real-person standing in front of the camera/display mirror, and allows the user to change clothes and other appearance attributes. Using Microsoft's Kinect, poses are matched to a video database of different torsos and legs, and "pages" showing different clothes are turned by hand gestures.

• Ph.D. Thesis 2014 Analyzing Tatonnement Dynamics in Economic Markets Cheung, Yun Kuen Abstract | PDF

  Title: Analyzing Tatonnement Dynamics in Economic Markets

  Candidate: Cheung, Yun Kuen

  Advisor(s): Cole, Richard

  Abstract:

  The impetus for this dissertation is to explain why well-functioning markets might be able to stay at or near a market equilibrium. We argue that tatonnement, a natural, simple and distributed price update dynamic in economic markets, is a plausible candidate to explain how markets might reach their equilibria.

  Tatonnement is broadly defined as follows: if the demand for a good is more than the supply, increase the price of the good, and conversely, decrease the price when the demand is less than the supply. Prior works show that tatonnement converges to market equilibrium in some markets while it fails to converge in other markets. Our goal is to extend the classes of markets in which tatonnement is shown to converge. The prior positive results largely concerned markets with substitute goods. We seek market constraints which enable tatonnement to converge in markets with complementary goods, or with a mixture of substitutes and complementary goods. We also show fast convergence rates for some of these markets.

  We introduce an amortized analysis technique to handle asynchronous events - in our case asynchronous price updates. On the other hand, for some markets we show that tatonnement is equivalent to generalized gradient descent (GGD). The amortized analysis and our analysis on GGD may be of independent interests.

• Ph.D. Thesis 2014 Low-latency Image Recognition withGPU-accelerated Convolutional Networksfor Web-based Services Huang, Fu Jie Abstract | PDF

  Title: Low-latency Image Recognition withGPU-accelerated Convolutional Networksfor Web-based Services

  Candidate: Huang, Fu Jie

  Advisor(s): LeCun, Yann

  Abstract:

  In this work, we describe an application of convolutional networks to object classification and detection in images. The task of image based object recognition is surveyed in the first chapter. Its application in internet advertisement is one of the main motivations of this work. 

  The architecture of the convolutional networks is described in details in the following chapter. Stochastic gradient descent is used to train the networks.
  
  We then describe the data collection and labelling process. The set of training data labelled basically decides what kind of recognizer is being built. Four binary classifers are trained for the object types of sailboat, car, motorbike, and dog.
  
  GPU based massive parallel implementation of the convolutional networks is built. This enables us to run the convolution operations at close to 40 times faster than running on a traditional CPU. Details about how to implement the convolutional operation on NVIDIA GPUs using CUDA is disscused.
  
  In order to apply the object recognizer in a production environment where millions of images are processed daily, we have built a platform with cloud computing. We describe how large scale and low latency image processing can be achieved with such a system. 

• Ph.D. Thesis 2014 Effective Algorithms for the Satisfiability of Quantifier-Free Formulas Over Linear Real and Integer Arithmetic King, Tim Abstract | PDF

  Title: Effective Algorithms for the Satisfiability of Quantifier-Free Formulas Over Linear Real and Integer Arithmetic

  Candidate: King, Tim

  Advisor(s): Barrett, Clark

  Abstract:

  A core technique of modern tools for formally reasoning about computing systems is generating and dispatching queries to automated theorem provers, including Satisfiability Modulo Theories (SMT) provers. SMT provers aim at the tight integration of decision procedures for propositional satisfiability and decision procedures for fixed first-order theories ‒ known as theory solvers. This thesis presents several advancements in the design and implementation of theory solvers for quantifier-free linear real, integer, and mixed integer and real arithmetic. These are implemented within the SMT system CVC4. We begin by formally describing the Satisfiability Modulo Theories problem and the role of theory solvers within CVC4. We discuss known techniques for building solvers for quantifier-free linear real, integer, and mixed integer and real arithmetic around the Simplex for DPLL(T) algorithm. We give several small improvements to theory solvers using this algorithm and describe the implementation and theory of this algorithm in detail. To extend the class of problems that the theory solver can robustly support, we borrow and adapt several techniques from linear programming (LP) and mixed integer programming (MIP) solvers which come from the tradition of optimization. We propose a new decicion procedure for quantifier-free linear real arithmetic that replaces the Simplex for DPLL(T) algorithm with a variant of the Simplex algorithm that performs a form of optimization ‒ minimizing the sum of infeasibilties. In this thesis, we additionally describe techniques for leveraging LP and MIP solvers to improve the performance of SMT solvers without compromising correctness. Previous efforts to leverage such solvers in the context of SMT have concluded that in addition to being potentially unsound, such solvers are too heavyweight to compete in the context of SMT. We present an empirical comparison against other state-of-the-art SMT tools to demonstrate the effectiveness of the proposed solutions.

• Ph.D. Thesis 2014 Cryptographic Algorithms for the SecureDelegation of Multiparty Computation Lopez-Alt, Adriana Abstract | PDF

  Title: Cryptographic Algorithms for the SecureDelegation of Multiparty Computation

  Candidate: Lopez-Alt, Adriana

  Advisor(s): Dodis, Yevgeniy

  Abstract:

  In today’s world, we store our data and perform expensive computations remotely on powerful servers (a.k.a. “the cloud”) rather than on our local devices. In this dissertation we study the question of achieving cryptographic security in the setting where multiple (mutually distrusting) clients wish to delegate the computation of a joint function on their inputs to an untrusted cloud, while keeping these inputs private. We introduce two frameworks for modeling such protocols.

  1. The first, called cloud-assisted multiparty computation (cloud-assisted MPC), builds on the standard notion of MPC to incorporate the concept of delegation. In particular, since the cloud is expected to perform the computation of the function, our definition requires the communication complexity of the protocol, as well as the computation time of all clients to be (essentially) independent of the complexity of the function.
  2. The second, called on-the-fly MPC, builds on the notion of cloud-assisted MPC and further requires that the clients be involved only when initially uploading their input to the cloud, and in a final phase when outputs are revealed. In particular, this allows the server to dynamically choose functions (and subsets of data on which to evaluate these functions) “on- the-fly”, and evaluate them without requiring any interaction with the clients. The only interaction required takes place in the final phase after the computation has been completed, when the clients must retroactively approve both the chosen functions, and the subsets of data upon which these functions were evaluated.

  We construct cloud-assisted and on-the-fly MPC protocols using fully homomorphic encryption (FHE). However, FHE requires inputs to be encrypted under the same key; we extend it to the multiparty setting in two ways:

  1. We introduce the notion of threshold FHE : fully homomorphic encryption that allows the clients to jointly generate a common public key (whose corresponding secret key is shared among them), as well as decrypt a ciphertext under this public key without learning any- thing but the plaintext. Using threshold FHE, we show how to construct an efficient cloud- assisted MPC protocol. We construct threshold FHE using (a modification of) the Brakerski- Vaikuntanathan (ring-based) FHE scheme; however our ideas extend to many other lattice- based FHE schemes in the literature.
  2. We introduce the notion of multikey FHE : fully homomorphic encryption that allows the cloud to perform homomorphic evaluation on ciphertexts encrypted under different and independent keys. We show a construction of on-the-fly MPC using multikey FHE, and construct a multikey FHE scheme based on NTRU encryption. We highlight that it was previously not known how to make NTRU fully homomorphic, even for a single key. Therefore, we view the construction of (multikey) FHE from NTRU encryption as a main contribution of independent interest.
• Ph.D. Thesis 2014 Robust and Efficient Methods for Approximation and Optimization of Stability Measures Mitchell, Tim Abstract | PDF

  Title: Robust and Efficient Methods for Approximation and Optimization of Stability Measures

  Candidate: Mitchell, Tim

  Advisor(s): Overton, Michael

  Abstract:

  We consider two new algorithms with practical application to the problem of designing controllers for linear dynamical systems with input and output: a new spectral value set based algorithm called hybrid expansion-contraction intended for approximating the H-infinity norm, or equivalently, the complex stability radius, of large-scale systems, and a new BFGS SQP based optimization method for nonsmooth, nonconvex constrained optimization motivated by multi-objective controller design. In comprehensive numerical experiments, we show that both algorithms in their respect domains are significantly faster and more robust compared to other available alternatives. Moreover, we present convergence guarantees for hybrid expansion-contraction, proving that it converges at least superlinearly, and observe that it converges quadratically in practice, and typically to good approximations to the H-infinity norm, for problems which we can verify this. We also extend the hybrid expansion-contraction algorithm to the real stability radius, a measure which is known to be more difficult to compute than the complex stability radius. Finally, for the purposes of comparing multiple optimization methods, we present a new visualization tool called relative minimization profiles that allow for simultaneously assessing the relative performance of algorithms with respect to three important performance characteristics, highlighting how these measures interrelate to one another and compare to the other competing algorithms on heterogenous test sets. We employ relative minimization profiles to empirically validate our proposed BFGS SQP method in terms of quality of minimization, attaining feasibility, and speed of progress compared to other available methods on challenging test sets comprised of nonsmooth, nonconvex constrained optimization problems arising in controller design.

• Ph.D. Thesis 2014 Building Efficient Distributed In-memory Systems Power, Russell Abstract | PDF

  Title: Building Efficient Distributed In-memory Systems

  Candidate: Power, Russell

  Advisor(s): Li, Jinyang

  Abstract:

  The recent cloud computing revolution has changed the distributed computing landscape, making the resources of entire datacenters available to ordinary users. This process has been greatly aided by dataflow style frameworks such as MapReduce which expose simple model for programs, allowing for efficient, fault-tolerant execution across many machines. While the MapReduce model has proved to be effective for many applications, there are a wide class of applications which are difficult to write or inefficient in such a model. This includes many familiar and important applications such as PageRank, matrix factorization and a number of machine learning algorithms. In lieu of a good framework for building these applications, users resort to writing applications using MPI or RPC, a difficult and error-prone construction.

  This thesis presents 2 complementary frameworks, Piccolo and Spartan, which help programmers to write in-memory distributed applications not served well by existing approaches.

  Piccolo presents a new data-centric programming model for in-memory applications. Unlike data-flow models, Piccolo allows programs running on different machines to share distributed, mutable state via a key-value table interface. This design allows for both high-performance and additional flexibility. Piccolo makes novel use of commutative updates to efficiently resolve write-write conflicts. We find Piccolo provides an efficient backend for a wide-range of applications: from PageRank and matrix multiplication to web-crawling.

  While Piccolo provides an efficient backend for distributed computation, it can still be some- what cumbersome to write programs using it directly. To address this, we created Spartan. Spartan implements a distributed implementation of the NumPy array language, and fully sup- ports important array language features such as spatial indexing (slicing), fancy indexing and broadcasting. A key feature of Spartan is its use of a small number of simple, powerful high-level operators to provide most functionality. Not only do these operators dramatically simplify the design and implementation of Spartan, they also allow users to implement new functionality with ease.

  We evaluate Piccolo and Spartan on a wide range of applications and find that they both perform significantly better than existing approaches.

• Ph.D. Thesis 2014 Runtime Compilation of Array-Oriented Python Programs Rubinsteyn, Alex Abstract | PDF

  Title: Runtime Compilation of Array-Oriented Python Programs

  Candidate: Rubinsteyn, Alex

  Advisor(s): Shasha, Dennis

  Abstract:

  The Python programming language has become a popular platform for data analysis and scientific computing. To mitigate the poor performance of Python's standard interpreter, numerically intensive computations are typically offloaded to library functions written in languages such as Fortran or C. If, however, some algorithm does not have an existing low-level implementation, then the scientific programmer must either accept sub-standard performance (sometimes orders of magnitude slower than native code) or themselves implement the desired functionality in a less productive but more efficient language.

  To alleviate this problem, this thesis present Parakeet, a runtime compiler for an array-oriented subset of Python. Parakeet does not replace the Python interpreter, but rather selectively augments it by compiling and executing functions explicitly marked by the programmer. Parakeet uses runtime type specialization to eliminate the performance-defeating dynamicism of untyped Python code. Parakeet's pervasive use of data parallel operators as a means for implementing array operations enables high-level restructuring optimization and compilation to parallel hardware such as multi-core CPUs and graphics processors. We evaluate Parakeet on a collection of numerical benchmarks and demonstrate its dramatic capacity for accelerating array-oriented Python programs.

   

• Ph.D. Thesis 2014 A Deep Learning Pipeline for Image Understanding and Acoustic Modeling Sermanet, Pierre Abstract | PDF

  Title: A Deep Learning Pipeline for Image Understanding and Acoustic Modeling

  Candidate: Sermanet, Pierre

  Advisor(s): LeCun, Yann

  Abstract:

  One of the biggest challenges artificial intelligence faces is making sense of the real world through sensory signals such as audio or video. Noisy inputs, varying object viewpoints, deformations and lighting conditions turn it into a high-dimensional problem which cannot be efficiently solved without learning from data.

  This thesis explores a general way of learning from high dimensional data (video, images, audio, text, financial data, etc.) called deep learning. It strives on the increasingly large amounts of data available to learn robust and invariant internal features in a hierarchical manner directly from the raw signals.

  We propose an unified pipeline for feature learning, recognition, localization and detection using Convolutional Networks (ConvNets) that can obtain state-of-the-art accuracy on a number of pattern recognition tasks, including acoustic modeling for speech recognition and object recognition in computer vision. ConvNets are particularly well suited for learning from continuous signals in terms of both accuracy and efficiency.

  Additionally, a novel and general deep learning approach to detection is proposed and successfully demonstrated on the most challenging vision datasets. We then generalize it to other modalities such as speech data. This approach allows accurate localization and detection objects in images or phones in voice signals by learning to predict boundaries from internal representations. We extend the reach of deep learning from classification to detection tasks in an integrated fashion by learning multiple tasks using a single deep model. This work is among the first to outperform human vision and establishes a new state of the art on some computer vision and speech recognition benchmarks.

   

• Ph.D. Thesis 2014 Towards New Interfaces For Pedagogy Stein, Murphy Abstract | PDF

  Title: Towards New Interfaces For Pedagogy

  Candidate: Stein, Murphy

  Advisor(s): Perlin, Ken

  Abstract:

  Developing technology to help people teach and learn is an important topic in Human Computer Interaction (HCI).

  In this thesis we present three studies on this topic. In the first study, we demonstrate new games for learning mathematics and discuss the evidence for key design decisions from user studies. In the second study, we develop a real-time video compositing system for distance education and share evidence for its potential value compared to standard techniques from two user studies. In the third study, we demonstrate our markerless hand tracking interface for real-time 3D manipulation and explain its advantages compared to other state-of-the-art methods.

  A data-driven methodology is applied intensively throughout the course of this study. Several paraphrase corpora are constructed using automatic techniques, experts and crowdsourcing platforms. Paraphrase systems are trained and evaluated by using these data as a cornerstone. We show that even with a very noisy or a relatively small amount of parallel training data, it is possible to learn paraphrase models which capture linguistic phenomena. This work expands the scope of paraphrase studies to targeting different language variations, and more potential applications, such as text normalization and domain adaptation.

• Ph.D. Thesis 2014 Computational Complexity Implicationsof Secure Coin-Flipping Tentes, Aristeidis Abstract | PDF

  Title: Computational Complexity Implicationsof Secure Coin-Flipping

  Candidate: Tentes, Aristeidis

  Advisor(s): Dodis, Yevgeniy

  Abstract:

  Modern Cryptography is based on computational intractability assumptions, e.g., Factoring, Discrete Logarithm, Diffie-Helman etc. However, since an assumption might be proven incorrect, there has been a lot of focus in order to construct cryptographic primitives based on the possibly most minimal assumption. The most popular minimal assumption, which is implied by the existence of almost all cryptographic primitives, is the existence of One Way Functions. Coin-Flipping protocols are known to be implied by One-Way Functions, however, a complete characterization of the inverse direction is not known. There was even speculation that weak notions of Coin Flipping Protocols might be strictly weaker than One Way Functions. In this thesis we show that even very weak notions of Coin Flipping protocols do imply One Way Functions. In particular we show that the existence of a coin-flipping protocol safe against any non-trivial constant bias (e.g 0.499) implies the existence of One Way Functions. This improves upon a recent result of Haitner and Omri [FOCS '11], who proved this implication for protocols with bias 0.207. Unlike the former result, our result also holds for weak coin-flipping protocols.

• Ph.D. Thesis 2014 Data-driven Approaches for Paraphrasing across Language Variations Xu, Wei Abstract | PDF

  Title: Data-driven Approaches for Paraphrasing across Language Variations

  Candidate: Xu, Wei

  Advisor(s): Grishman, Ralph

  Abstract:

  Our language changes very rapidly, accompanying political, social and cultural trends, as well as the evolution of science and technology. The Internet, especially the social media, has accelerated this process of change. This poses a severe challenge for both human beings and natural language processing (NLP) systems, which usually only model a snapshot of language presented in the form of text corpora within a certain domain and time frame.

  While much previous effort has investigated monolingual paraphrase and bilingual translation, we focus on modeling meaning-preserving transformations between variants of a single language. We use Shakespearean and Internet language as examples to investigate various aspects of this new paraphrase problem, including acquisition, generation, detection and evaluation.

  A data-driven methodology is applied intensively throughout the course of this study. Several paraphrase corpora are constructed using automatic techniques, experts and crowdsourcing platforms. Paraphrase systems are trained and evaluated by using these data as a cornerstone. We show that even with a very noisy or a relatively small amount of parallel training data, it is possible to learn paraphrase models which capture linguistic phenomena. This work expands the scope of paraphrase studies to targeting different language variations, and more potential applications, such as text normalization and domain adaptation.

• Ph.D. Thesis 2014 Positive-Unlabeled Learning in the Context of Protein Function Prediction Youngs, Noah Abstract | PDF

  Title: Positive-Unlabeled Learning in the Context of Protein Function Prediction

  Candidate: Youngs, Noah

  Advisor(s): Shasha, Dennis

  Abstract:

  With the recent proliferation of large, unlabeled data sets, a particular subclass of semisupervised learning problems has become more prevalent. Known as positiveunlabeled learning (PU learning), this scenario provides only positive labeled examples, usually just a small fraction of the entire dataset, with the remaining examples unknown and thus potentially belonging to either the positive or negative class. Since the vast majority of traditional machine learning classifiers require both positive and negative examples in the training set, a new class of algorithms has been developed to deal with PU learning problems.

  A canonical example of this scenario is topic labeling of a large corpus of documents. Once the size of a corpus reaches into the thousands, it becomes largely infeasible to have a curator read even a sizable fraction of the documents, and annotate them with topics. In addition, the entire set of topics may not be known, or may change over time, making it impossible for a curator to annotate which documents are NOT about certain topics. Thus a machine learning algorithm needs to be able to learn from a small set of positive examples, without knowledge of the negative class, and knowing that the unlabeled training examples may contain an arbitrary number of additional but as yet unknown positive examples. Another example of a PU learning scenario recently garnering attention is the protein function prediction problem (PFP problem).

  While the number of organisms with fully sequenced genomes continues to grow, the progress of annotating those sequences with the biological functions that they perform lags far behind. Machine learning methods have already been successfully applied to this problem, but with many organisms having a small number of positive annotated training examples, and the lack of availability of almost any labeled negative examples, PU learning algorithms can make large gains in predictive performance.

  The first part of this dissertation motivates the protein function prediction problem, explores previous work, and introduces novel methods that improve upon previously reported benchmarks for a particular type of learning algorithm, known as Gaussian Random Field Label Propagation (GRFLP). In addition, we present improvements to the computational efficiency of the GRFLP algorithm, and a modification to the traditional structure of the PFP learning problem that allows for simultaneous prediction across multiple species.

  The second part of the dissertation focuses specifically on the positive-unlabeled aspects of the PFP problem. Two novel algorithms are presented, and rigorously compared to existing PU learning techniques in the context of protein function prediction. Additionally, we take a step back and examine some of the theoretical considerations of the PU scenario in general, and provide an additional novel algorithm applicable in any PU context. This algorithm is tailored for situations in which the labeled positive examples are a small fraction of the set of true positive examples, and where the labeling process may be subject to some type of bias rather than being a random selection of true positives (arguably some of the most difficult PU learning scenarios).

  The third and fourth sections return to the PFP problem, examining the power of tertiary structure as a predictor of protein function, as well as presenting two case studies of function prediction performance on novel benchmarks. Lastly, we conclude with several promising avenues of future research into both PU learning in general, and the protein function prediction problem specifically.

• Ph.D. Thesis 2014 Hierarchical Convolutional Deep Learning in Computer Vision Zeiler, Matthew Abstract | PDF

  Title: Hierarchical Convolutional Deep Learning in Computer Vision

  Candidate: Zeiler, Matthew

  Advisor(s): Fergus, Rob

  Abstract:

  It has long been the goal in computer vision to learn a hierarchy of features useful for object recognition. Spanning the two traditional paradigms of machine learning, unsupervised and supervised learning, we investigate the application of deep learning methods to tackle this challenging task and to learn robust representations of images.

  We begin our investigation with the introduction of a novel unsupervised learning technique called deconvolutional networks. Based on convolutional sparse coding, we show this model learns interesting decompositions of images into parts without object label information. This method, which easily scales to large images, becomes increasingly invariant by learning multiple layers of feature extraction coupled with pooling layers. We introduce a novel pooling method called Gaussian pooling to enable these layers to store continuous location information while being differentiable, creating a unified objective function to optimize.

  In the supervised learning domain, a well-established model for recognition of objects is the convolutional network. We introduce a new regularization method for convolutional networks called stochastic pooling which relies on sampling noise to prevent these powerful models from overfitting. Additionally, we show novel visualizations of these complex models to better understand what they learn and to provide insight on how to develop state-of-the-art architectures for large-scale classification of 1,000 different object categories.

  We also investigate some other related problems in deep learning. First, we introduce a model for the task of mapping one high dimensional time series sequence onto another. Second, we address the choice of nonlinearity in neural networks, showing evidence that rectified linear units outperform others types in automatic speech recognition. Finally, we introduce a novel optimization method called ADADELTA which shows promising convergence speeds in practice while being robust to hyper-parameter selection.

• Ph.D. Thesis 2013 Incentive-Centered Design of Money-Free Mechanisms Gkatzelis, Vasilis Abstract | PDF

  Title: Incentive-Centered Design of Money-Free Mechanisms

  Candidate: Gkatzelis, Vasilis

  Advisor(s): Cole, Richard

  Abstract:

  This thesis serves as a step toward a better understanding of how to design fair and efficient multiagent resource allocation systems by bringing the incentives of the participating agents to the center of the design process. As the quality of these systems critically depends on the ways in which the participants interact with each other and with the system, an ill-designed set of incentives can lead to severe inefficiencies. The special focus of this work is on the problems that arise when the use of monetary exchanges between the system and the participants is prohibited. This is a common restriction that substantially complicates the designer's task; we nevertheless provide a sequence of positive results in the form of mechanisms that maximize efficiency or fairness despite the possibly self-interested behavior of the participating agents.

  The first part of this work is a contribution to the literature on approximate mechanism design without money. Given a set of divisible resources, our goal is to design a mechanism that allocates them among the agents. The main complication here is due to the fact that the agents' preferences over different allocations of these resources may not be known to the system. Therefore, the mechanism needs to be designed in such a way that it is in the best interest of every agent to report the truth about her preferences; since monetary rewards and penalties cannot be used in order to elicit the truth, a much more delicate regulation of the resource allocation is necessary. Our contribution mostly revolves around a new truthful mechanism that we propose, which we call the /Partial Allocation/ mechanism. We first show how to use the two-agent version of this mechanism to create a system with the best currently known worst-case efficiency guarantees for problem instances involving two agents. We then consider fairness measures and prove that the general version of this elegant mechanism yields surprisingly good approximation guarantees for the classic problem of fair division. More specifically, we use the well established solution of /Proportional Fairness/ as a benchmark and we show that for an arbitrary number of agents and resources, and for a very large class of agent preferences, our mechanism provides /every agent/ with a value close to her proportionally fair value. We complement these results by also studying the limits of truthful money-free mechanisms, and by providing other mechanisms for special classes of problem instances. Finally, we uncover interesting connections between our mechanism and the Vickrey-Clarke-Groves mechanism from the literature on mechanism design with money.

  The second part of this work concerns the design of money-free resource allocation mechanisms for /decentralized/ multiagent systems. As the world has become increasingly interconnected, such systems are using more and more resources that are geographically dispersed; in order to provide scalability in these systems, the mechanisms need to be decentralized. That is, the allocation decisions for any given resource should not assume global information regarding the system's resources or participants. We approach this restriction by using /coordination mechanisms/: a collection of simple resource allocation policies, each of which controls only one of the resources and uses only local information regarding the state of the system. The system's participants, facing these policies, have the option of choosing which resources they will access. We study a variety of coordination mechanisms and we prove that the social welfare of any equilibrium of the games that these mechanisms induce is a good approximation of the optimal welfare. Once again, we complement our positive results by studying the limits of coordination mechanisms. We also provide a detailed explanation of the seemingly counter-intuitive incentives that some of these mechanisms yield. Finally, we use this understanding in order to design a combinatorial constant-factor approximation algorithm for maximizing the social welfare, thus providing evidence that a game-theoretic mindset can lead to novel optimization algorithms.

• Ph.D. Thesis 2013 Locality Optimization for Data Parallel Programs Hielscher, Eric Abstract | PDF

  Title: Locality Optimization for Data Parallel Programs

  Candidate: Hielscher, Eric

  Advisor(s): Shasha, Dennis

  Abstract:

  Productivity languages such as NumPy and Matlab make it much easier to implement data-intensive numerical algorithms than it is to implement them in efficiency languages such as C++. This is important as many programmers (1) aren't expert programmers; or (2) don't have time to tune their software for performance, as their main job focus is not programming per se. The tradeoff is typically one of execution time versus programming time, as unless there are specialized library functions or precompiled primitives for your particular task a productivity language is likely to be orders of magnitude slower than an efficiency language.

  In this thesis, we present Parakeet, an array-oriented language embedded within Python, a widely-used productivity language. The Parakeet just-in-time compiler dynamically translates whole user functions to high performance multi-threaded native code. This thesis focuses in particular on our use of data parallel operators as a basis for locality enhancing program optimizations. e transform Parakeet programs written with the classic data parallel operators (Map, Reduce, and Scan; in Parakeet these are called adverbs) to process small local pieces (called tiles) of data at a time. To express this locality we introduce three new adverbs: TiledMap, TiledReduce, and TiledScan. These tiled adverbs are not exposed to the programmer but rather are automatically generated by a tiling transformation.

  We use this tiling algorithm to bring two classic locality optimizations to a data parallel setting: cache tiling, and register tiling. We set register tile sizes statically at compile time, but use an online autotuning search to find good cache tile sizes at runtime. We evaluate Parakeet and these optimizations on various benchmark programs, and exhibit excellent performance even compared to typical C implementations.

• Ph.D. Thesis 2013 Piecewise Smooth Surfaces with Features Kovacs, Denis Abstract | PDF

  Title: Piecewise Smooth Surfaces with Features

  Candidate: Kovacs, Denis

  Advisor(s): Zorin, Denis

  Abstract:

  The creation, manipulation and display of piecewise smooth surfaces has been a fundamental topic in computer graphics since its inception. The applications range from highest-quality surfaces for manufacturing in CAD, to believable animations of virtual creatures in Special Effects, to virtual worlds rendered in real-time in computer games. 

  Our focus is on improving the a) mathematical representation and b) automatic construction of such surfaces from finely sampled meshes in the presence of features. Features can be areas of higher geometric detail in an otherwise smooth area of the mesh, or sharp creases that contrast the overall smooth appearance of an object.
  In the first part, we build on techniques that define piecewise smooth surfaces, to improve their quality in the presence of features. We present a crease technique suitable for real-time applications that helps increases the perceived visual detail of objects that are required to be very compactly represented and efficiently evaluated.
  We then introduce a new subdivision scheme that allows the use of T-junctions for better local refinement. It thus reduces the need for extraordinary vertices, which can cause surface artifacts especially on animated objects.
  In the second part, we consider the problem of how to build the control meshes of piecewise smooth surfaces, in a way that the resulting surface closely approximates an existing data set (such as a 3D range scan), particularly in the presence of features. To this end, we introduce a simple modification that can be applied to a wide range of parameterization techniques to obtain an anisotropic parameterization. We show that a resulting quadrangulation can indeed better approximate the original surface. Finally, we present a quadrangulation scheme that turns a data set into a quad mesh with T-junctions, which we then use as a T-Spline control mesh to obtain a smooth surface.

• Ph.D. Thesis 2013 Low-level Image Priors and Laplacian Preconditioners for Applications in Computer Graphics and Computational Photography Krishnan, Dilip Abstract | PDF

  Title: Low-level Image Priors and Laplacian Preconditioners for Applications in Computer Graphics and Computational Photography

  Candidate: Krishnan, Dilip

  Advisor(s): Fergus, Rob

  Abstract:

  In the first part of this thesis, we develop novel image priors and efficient algorithms for image denoising and deconvolution applications. Our priors and algorithms enable fast, high-quality restoration of images corrupted by noise or blur. In the second part, we develop effective preconditioners for Laplacian matrices. Such matrices arise in a number of computer graphics and computational photography problems such as image colorization, tone mapping and geodesic distance computation on 3D meshes.

  The first prior we develop is a spectral prior which models correlations between different spectral bands. We introduce a prototype camera and flash system, used in conjunction with the spectral prior, to enable taking photographs at very low light levels. Our second prior is a sparsity-based measure for blind image deconvolution. This prior gives lower costs to sharp images than blurred ones, enabling the use simple and efficient Maximum a-Posteriori algorithms.

  We develop a new algorithm for the non-blind deconvolution problem. This enables extremely fast deconvolution of images blurred by a known blur kernel. Our algorithm uses Fast Fourier Transforms and Lookup Tables to achieve real-time deconvolution performance with non convex gradient-based priors. Finally, for certain image restoration problems with no clear formation model, we demonstrate how learning a direct mapping between original/corrupted patch pairs enables effective restoration.

  We develop multi-level preconditioners to solve discrete Poisson equations. Existing multilevel preconditioners have two major drawbacks: excessive bandwidth growth at coarse levels; and the inability to adapt to problems with highly varying coefficients. Our approach tackles both these problems by introducing sparsification and compensation steps at each level. We interleave the selection of fine and coarse-level variables with the removal of weak connections between potential fine-level variables (sparsification) and compensate for these changes by strengthening nearby connections. By applying these operations before each elimination step and repeating the procedure recursively on the resulting smaller systems, we obtain highly efficient schemes. The construction is linear in time and memory. Numerical experiments demonstrate that our new schemes outperform state of the art methods, both in terms of operation count and wall-clock time, over a range of 2D and 3D problems.

• Ph.D. Thesis 2013 Relation Extraction with Weak Supervision and Distributional Semantics Min, Bonan Abstract | PDF

  Title: Relation Extraction with Weak Supervision and Distributional Semantics

  Candidate: Min, Bonan

  Advisor(s): Grishman, Ralph

  Abstract:

  Relation Extraction aims at detecting and categorizing semantic relations between pairs of entities in unstructured text. It benefits an enormous number of applications such as Web search and Question Answering. Traditional approaches for relation extraction either rely on learning from a large number of accurate human-labeled examples or pattern matching with hand-crafted rules. These resources are very laborious to obtain and can only be applied to a narrow set of target types of interest.

  This talk focuses on learning relations with little or no human supervision. First, we examine the approach that treats relation extraction as a supervised learning problem. We develop an algorithm that is able to train a model with approximately 1/3 of the human-annotation cost and that matches the performance of models trained with high-quality annotation. Second, we investigate distant supervision, a weakly supervised algorithm that automatically generates its own labeled training data. We develop a latent Bayesian framework for this purpose. By using a model which provides a better approximation of the weak source of supervision, it outperforms the state-of-the-art methods. Finally, we investigate the possibility of building all relational tables beforehand with an unsupervised relation extraction algorithm. We develop an effective yet efficient algorithm that combines the power of various semantic resources that are automatically mined from a corpus based on distributional semantics. The algorithm is able to extract a very large set of relations from the web at high precision.

• Ph.D. Thesis 2013 Usable Security Mechanisms in the Developing World Paik, Michael Abstract | PDF

  Title: Usable Security Mechanisms in the Developing World

  Candidate: Paik, Michael

  Advisor(s): Subramanian; Lakshminarayanan

  Abstract:

  Security and privacy are increasingly important in our interconnected world. Cybercrimes, including identity theft, phishing, and other attacks, are on the rise, and computer-assisted crimes such as theft and stalking are becoming commonplace.

  Contemporary with this trend is the uptake of technology in the developing world, proceeding at a pace often outstripping that of the developed world. Penetration of mobile phones and services such as healthcare delivery, mobile money, and social networking is higher than that of even amenities like electricity. Connectivity is empowering disenfranchised people, providing information and services to the heretofore disconnected poor.

  There are efforts to use technology to enhance physical security and well-being in the developing world, including citizen journalism, education, improving drug security, attendance tracking, etc.

  However, there are significant challenges to security both in the digital and the physical domains that are particular to these contexts. Infrastructure is constrained, literacy, numeracy, and familiarity with basic technologies cannot be assumed, and environments are harsh on hardware. These circumstances often prevent security best practices from being transplanted directly to these regions – in many ways, the adoption of technology has overtaken the users ability to use it safely, and their trust in it is oftentimes reater than it should be.

  This dissertation describes several systems and methodologies designed to operate in the developing world, using technologies and metaphors that are familiar to users and that are robust against the operating environments.

  It begins with an overview of the state of affairs, and several threat models. It continues with a description of Signet, a method to use SIM cards as trusted computing hardware to provide secure signed receipts. Next, Epothecary describes a low-infrastructure system for tracking pharmaceuticals that also significantly and asymmetrically increases costs for counterfeiters. The balance consists of a description of a low-cost Biometric Terminal currently in use by NGOs in India performing DOTS-based tuberculosis treatment, Blacknoise, an investigation into the use of low-cost cameraphones with noisy imaging sensors for image-based steganography, and finally Innoculous, a low-cost, crowdsourcing system for combating the spread of computer viruses, particularly among non-networked computers, while also collecting valuable "epidemiological" data.

• Ph.D. Thesis 2013 Inapproximability Reductions and Integrality Gaps Popat, Preyas Abstract | PDF

  Title: Inapproximability Reductions and Integrality Gaps

  Candidate: Popat, Preyas

  Advisor(s): Khot, Subhash

  Abstract:

  In this thesis we prove intractability results for several well studied problems in combinatorial optimization.

  Closest Vector Problem with Preprocessing (CVPP): We show that the preprocessing version of the well known Closest Vector Problem is hard to approximate to an almost polynomial factor unless NP is in quasi polynomial time. The approximability of CVPP is closely related to the security of lattice based cryptosystems.

  Pricing Loss Leaders: We show hardness of approximation results for the problem of maximizing profit from buyers with single minded valuations where each buyer is interested in bundles of at most k items, and the items are allowed to have negative prices ("Loss Leaders"). For k = 2, we show that assuming the Unique Games Conjecture, it is hard to approximate the profit to any constant factor. For k > 2, we show the same result assuming P != N P.

  Integrality gaps: We show SemiDefinite Programming (SDP) integrality gaps for Unique Games and 2 to 1 Games. Inapproximability results for these problems imply inapproximability results for many fundamental optimization problems. For the first problem, we show "approximate" integrality gaps for super constant rounds of the powerful Lasserre hierarchy. For the second problem we show integrality gaps for the basic SDP relaxation with perfect completeness.

• Ph.D. Thesis 2013 Natural Interaction with a Virtual World Rosenberg, Ilya Abstract | PDF

  Title: Natural Interaction with a Virtual World

  Candidate: Rosenberg, Ilya

  Advisor(s): Perlin, Ken

  Abstract:

  A large portion of computer graphics and human/computer interaction is concerned with the creation, manipulation and use of two and three dimensional objects existing in a virtual world. By creating more natural physical interfaces and virtual worlds which behave in physically plausible ways, it is possible to empower nonexpert users to create, work and play in virtual environments. This thesis is concerned with the design, creation, and optimization of user-input devices which break down the barriers between the real and the virtual as well as the development of software algorithms which allow for the creation of physically realistic virtual worlds.

• Ph.D. Thesis 2013 Security Mechanisms for Physical Authentication Sharma, Ashlesh Abstract | PDF

  Title: Security Mechanisms for Physical Authentication

  Candidate: Sharma, Ashlesh

  Advisor(s): Subramanian; Lakshminarayanan

  Abstract:

  Counterfeiting of goods is a worldwide problem where the losses are in billions of dollars. It is estimated that 10% of all the world trade is counterfeit. To alleviate counterfeiting, a number of techniques are used from barcodes to holograms. But these technologies are easily reproducible and hence they are ineffective against counterfeiters.

  In this thesis, we introduce PaperSpeckle, a novel way to fingerprint any piece of paper based on its unique microscopic properties. Next, we extend and generalize this work to introduce TextureSpeckle, a novel way to fingerprint and characterize the uniqueness of the surface of a material based on the interaction of light with the natural randomness present in the rough structure at the microscopic level of the surface. We show the existence and uniqueness of these fingerprints by analyzing a large number of surfaces (over 20,000 microscopic surfaces and 200 million pairwise comparisons) of different materials. We also define the entropy of the fingerprints and show how each surface can be uniquely identified in a robust manner even in case of damage.

  From a theoretical perspective, we consider a discrete approximation model from light scattering theory which allows us to compute the speckle pattern for a given surface. Under this computational model, we show that given a speckle pattern, it is computationally hard to reconstruct the physical surface characteristics by simulating the multiple scattering of light. Using TextureSpeckle as a security primitive, we design secure protocols to enable a variety of scenarios such as: i) supply chain security, where applications range from drug tracking to inventory management, ii) mobile based secure transfer of money (mobile money), where any paper can be changed to an on-demand currency, and iii) fingerprint ecosystem, a cloud based system, where any physical object can be identified and authenticated on-demand.

  We discuss the construction of the prototype device ranging from optical lens design to usability aspects and show how our technique can be applied in the real world to alleviate counterfeiting and forgery. In addition, we introduce Pattern Matching Puzzles (PMPs), a usable security mechanism that provides a 'human computable' one-time-MAC (message authentication code) for every transaction,making each transaction information-theoretically secure against various adversarial attacks. The puzzles are easy tosolve even for semi-literate users with simple pattern recognition skills.

• Ph.D. Thesis 2013 Augmenting Information Flow for Visual Privacy Spiro, Ian Abstract | PDF

  Title: Augmenting Information Flow for Visual Privacy

  Candidate: Spiro, Ian

  Advisor(s): Bregler, Christopher

  Abstract:

  In the Information Age, visual media take on powerful new forms. Photographs once printed on paper and stored in physical albums now exist as digital files. With the rise of social media, photo data has moved to the cloud for rapid dissemination. The upside can be measured in terms of increased efficiency, greater reach, or reduced printing costs. But there is a downside that is harder to quantify: the risk of private photos or videos leaking inappropriately. Human imagery is potentially sensitive, revealing private details of a person’s body, lifestyle, activities, and more. Images create visceral responses and have the potential to permanently damage a person’s reputation.

  We employed the theory of contextual integrity to explore privacy aspects of transmitting the human form. In response to privacy threats from new sociotechnical systems, we developed practical solutions that have the potential to restore balance. The main work is a set of client-side, technical interventions that can be used to alter information flows and provide features to support visual privacy. In the first approach, we use crowdsourcing to extract specific, useful human signal from video to decouple it from bundled identity information. The second approach is an attempt to achieve similar ends with pure software. Instead of using information workers, we developed a series of filters that alter video to hide identity information while still revealing motion signal. The final approach is an attempt to control the recipients of photos by encoding them in the visual channel. The software completely protects data from third-parties who lack proper credentials and maintains data integrity by exploiting the visual coherence of uploaded images, even in the face of JPEG compression. The software offers end-to-end encryption that is compatible with existing social media applications.

• Ph.D. Thesis 2013 Toward a computational solution to the inverse problem of how hypoxia arises in metabolically heterogeneous cancer cell populations Sundstrom, Andrew Abstract | PDF

  Title: Toward a computational solution to the inverse problem of how hypoxia arises in metabolically heterogeneous cancer cell populations

  Candidate: Sundstrom, Andrew

  Advisor(s): Mishra, Bud; Bar-Sagi, Dafna

  Abstract:

  As a tumor grows, it rapidly outstrips its blood supply, leaving portions of tumor that undergo hypoxia. Hypoxia is strongly correlated with poor prognosis as it renders tumors less responsive to chemotherapy and radiotherapy. During hypoxia, HIFs upregulate production of glycolysis enzymes and VEGF, thereby promoting metabolic heterogeneity and angiogenesis, and proving to be directly instrumental in tumor progression. Prolonged hypoxia leads to necrosis, which in turn activates inflammatory responses that produce cytokines that stimulate tumor growth. Hypoxic tumor cells interact with macrophages and fibroblasts, both involved with inflammatory processes tied to tumor progression. So it is of clinical and theoretical significance to understand: Under what conditions does hypoxia arise in a heterogeneous cell population? Our aim is to transform this biological origins problem into a computational inverse problem, and then attack it using approaches from computer science. First, we develop a minimal, stochastic, spatiotemporal simulation of large heterogeneous cell populations interacting in three dimensions. The simulation can manifest stable localized regions of hypoxia. Second, we employ and develop a variety of algorithms to analyze histological images of hypoxia in xenographed colorectal tumors, and extract features to construct a spatiotemporal logical characterization of hypoxia. We also consider characterizing hypoxia by a linear regression functional learning mechanism that yields a similarity score. Third, we employ a Bayesian statistical model checking algorithm that can determine, over some bounded number of simulation executions, whether hypoxia is likely to emerge under some fixed set of simulation parameters, and some fixed logical or functional description of hypoxia. Driving the model checking process is one of three adaptive Monte Carlo sampling algorithms we developed to explore the high dimensional space of simulation initial conditions and operational parameters. Taken together, these three system components formulate a novel approach to the inverse problem above, and constitute a design for a tool that can be placed into the hands of experimentalists, for testing hypotheses based upon known parameter values or ones the tool might discover. In principle, this design can be generalized to other biological phenomena involving large heterogeneous populations of interacting cells.

• Ph.D. Thesis 2013 Rethinking Information Privacy for the Web Tierney, Matthew Abstract | PDF

  Title: Rethinking Information Privacy for the Web

  Candidate: Tierney, Matthew

  Advisor(s): Subramanian; Lakshminarayanan

  Abstract:

  In response to Supreme Court Justice Samuel Alito’s opinion that society should accept a decline in personal privacy with modern technology, Hanni M. Fakhoury, staff attorney with the Electronic Frontier Foundation, argued “Technology doesn’t involve an ‘inevitable’ tradeoff [of increased convenience] with privacy. The only inevitability must be the demand that privacy be a value built into our technology” [42]. Our position resonates with Mr. Fakhoury’s. In this thesis, we present three artifacts that address the balance between usability, efficiency, and privacy as we rethink information privacy for the web.

  In the first part of this thesis, we present the design, implementation and evaluation of Cryptagram, a system designed to enhance online photo privacy. Cryptagram enables users to convert photos into encrypted images, which the users upload to Online Social Networks (OSNs). Users directly manage access control to those photos via shared keys that are independent of OSNs or other third parties. OSNs apply standard image transformations (JPEG compression) to all uploaded images so Cryptagram provides image encoding and encryption protocols that are tolerant to these transformations. Cryptagram guarantees that the recipient with the right credentials can completely retrieve the original image from the transformed version of the uploaded encrypted image while the OSN cannot infer the original image. Cryptagram’s browser extension integrates seamlessly with preexisting OSNs, including Facebook and Google+, and currently has over 400 active users.

  In the second part of this thesis, we present the design and implementation of Lockbox, a system designed to provide end-to-end private file-sharing with the convenience of Google Drive or Dropbox. Lockbox uniquely combines two important design points: (1) a federated system for detecting and recovering from server equivocation and (2) a hybrid cryptosystem over delta encoded data to balance storage and bandwidth costs with efficiency for syncing end-user data. To facilitate appropriate use of public keys in the hybrid cryptosystem, we integrate a service that we call KeyNet, which is a web service designed to leverage existing authentication media (e.g., OAuth, verified email addresses) to improve the usability of public key cryptography.

  In the third part of this thesis, we present the design of Compass, which realizes the philosophical privacy framework of contextual integrity (CI) as a full OSN design. CI), which we believe better captures users privacy expectations in OSNs. In Compass, three properties hold: (a) users are associated with roles in specific contexts; (b) every piece of information posted by a user is associated with a specific context; (c) norms defined on roles and attributes of posts in a context govern how information is shared across users within that context. Given the definition of a context and its corresponding norm set, we describe the design of a compiler that converts the human-readable norm definitions to generate appropriate information flow verification logic including: (a) a compact binary decision diagram for the norm set; and (b) access control code that evaluates how a new post to a context will flow. We have implemented a prototype that shows how the philosophical framework of contextual integrity can be realized in practice to achieve strong privacy guarantees with limited additional verification overhead.

• Ph.D. Thesis 2012 Learning Hierarchical Feature Extractors For ImageRecognition Boureau, Y-Lan Abstract | PDF

  Title: Learning Hierarchical Feature Extractors For ImageRecognition

  Candidate: Boureau, Y-Lan

  Advisor(s): LeCun, Yann

  Abstract:

  Telling cow from sheep is effortless for most animals, but requires much engineering for computers. In this thesis, we seek to tease out basic principles that underlie many recent advances in image recognition. First, we recast many methods into a common unsupervised feature extraction framework based on an alternation of coding steps, which encode the input by comparing it with a collection of reference patterns, and pooling steps, which compute an aggregation statistic summarizing the codes within some region of interest of the image.

  Within that framework, we conduct extensive comparative evaluations of many coding or pooling operators proposed in the literature. Our results demonstrate a robust superiority of sparse coding (which decomposes an input as a linear combination of a few visual words) and max pooling (which summarizes a set of inputs by their maximum value). We also propose macrofeatures, which import into the popular spatial pyramid framework the joint encoding of nearby features commonly practiced in neural networks, and obtain significantly improved image recognition performance. Next, we analyze the statistical properties of max pooling that underlie its better performance, through a simple theoretical model of feature activation. We then present results of experiments that confirm many predictions of the model. Beyond the pooling operator itself, an important parameter is the set of pools over which the summary statistic is computed. We propose locality in feature configuration space as a natural criterion for devising better pools. Finally, we propose ways to make coding faster and more powerful through fast convolutional feedforward architectures, and examine how to incorporate supervision into feature extraction schemes. Overall, our experiments offer insights into what makes current systems work so well, and state-of-the-art results on several image recognition benchmarks.

• Ph.D. Thesis 2012 On populations, haplotypes and genome sequencing Franquin, Pierre Abstract | PDF

  Title: On populations, haplotypes and genome sequencing

  Candidate: Franquin, Pierre

  Advisor(s): Mishra, Bud

  Abstract:

  Population genetics has seen a renewed interest since the completion of the human genome project. With the availability of rapidly growing volumes of genomic data, the scientific and medical communities have been optimistic that better understanding of human diseases as well as their treatment were imminent. Many population genomic models and association studies have been designed (or redesigned) to address these problems. For instance, the genome-wide association studies (GWAS) had raised hopes for finding disease markers, personalized medicine and rational drug design. Yet, as of today, they have not yielded results that live up to their promise and have only led to a frustrating disappointment.

  Intrigued, but not deterred by these challenges, this dissertation visits the different aspects of these problems. In the first part, we will review the different models and theories of population genetics that are now challenged. We will propose our own implementation of a model to test different hypotheses. This effort will hopefully help us in understanding whether our expectations were unreasonably too high or if we had ignored a crucial piece of information. When discussing association studies, we must not forget that we rely on data that are produced by sequencing technologies, so far available. We have to ensure that the quality of this data is reasonably good for GWAS. Unfortunately, as we will see in the second part, despite the existence of a diverse set of sequencing technologies, none of them can produce haplotypes with phasing, which appears to be the most important type of sequence data needed for association studies. To address this challenge, we propose a novel approach for a sequencing technology, called SMASH that allows us to create the quality and type of haplotypic genome sequences necessary for efficient population genetics.

• Ph.D. Thesis 2012 Optimizing Machine Translation by Learning to Search Galron, Daniel Abstract | PDF

  Title: Optimizing Machine Translation by Learning to Search

  Candidate: Galron, Daniel

  Advisor(s): Melamed, Dan

  Abstract:

  We present a novel approach to training discriminative tree-structured machine translation systems by learning to search. We describe three primary innovations in this work: a new parsing coordinator architecture and algorithms to synthesize the required training examples for the learning algorithm; a new semiring that provides an unbiased way to compare translations; and a new training objective that measures whether a translation inference improves the quality of a translation. We also apply the reinforcement learning concept of exploration to SMT. Finally, we empirically evaluate the effects of our innovations on the quality of translations output by our system.

• Ph.D. Thesis 2012 Flexible-Cost SLAM Grimes, Matthew Abstract | PDF

  Title: Flexible-Cost SLAM

  Candidate: Grimes, Matthew

  Advisor(s): LeCun, Yann

  Abstract:

  The ability of a robot to track its position and its surroundings is critical in mobile robotics applications, such as autonomous transport, farming, search-and-rescue, and planetary exploration.

  As a foundational building block to such tasks, localization must remain reliable and unobtrusive. For example, it must not provide an unneeded level of precision, when the cost of doing so displaces higher-level tasks from a busy CPU. Nor should it produce noisy estimates on the cheap, when there are CPU cycles to spare.

  This thesis explores localization solutions that provide exactly the amount of accuracy needed to a given task. We begin with a real-world system used in the DARPA Learning Applied to Ground Robotics (LAGR) competition. Using a novel hybrid of wheel and visual odometry, we cut the cost of visual odometry from 100% of a CPU to 5%, clearing room for other critical visual processes, such as long-range terrain classification. We present our hybrid odometer in chapter 2.

  Next, we describe a novel SLAM algorithm that provides a means to choose the desired balance between cost and accuracy. At its fastest setting, our algorithm converges faster than previous stochastic SLAM solvers, while maintaining significantly better accuracy. At its most accurate, it provides the same solution as exact SLAM solvers. Its main feature, however, is the ability to flexibly choose any point between these two extremes of speed and precision, as circumstances demand. As a result, we are able to guarantee real-time performance at each timestep on city-scale maps with large loops. We present this solver in chapter 3, along with results from both commonly available datasets and Google Street View data.

  Taken as a whole, this thesis recognizes that precision and efficiency can be competing values, whose proper balance depends on the application and its fluctuating circumstances. It demonstrates how a localizer can and should fit its cost to the task at hand, rather than the other way around. In enabling this flexibility, we demonstrate a new direction for SLAM research, as well as provide a new convenience for end-users, who may wish to map the world without stopping it.

• Ph.D. Thesis 2012 SMT Beyond DPLL(T): A New Approach to Theory Solvers and Theory Combination Jovanovic, Dejan Abstract | PDF

  Title: SMT Beyond DPLL(T): A New Approach to Theory Solvers and Theory Combination

  Candidate: Jovanovic, Dejan

  Advisor(s): Barrett, Clark

  Abstract:

  Satisifiability modulo theories (SMT) is the problem of deciding whether a given logical formula can be satisifed with respect to a combination of background theories. The past few decades have seen many significant developments in the field, including fast Boolean satisfiability solvers (SAT), efficient decision procedures for a growing number of expressive theories, and frameworks for modular combination of decision procedures. All these improvements, with addition of robust SMT solver implementations, culminated with the acceptance of SMT as a standard tool in the fields of automated reasoning and computer aided verification. In this thesis we develop new decision procedures for the theory of linear integer arithmetic and the theory of non-linear real arithmetic, and develop a new general framework fro combination of decision procedures. The new decision procedures integrate theory specific reasoning and the Boolean search to provide more powerful and efficient procedures, and allow a more expressive language for explaining problematic states. The new framework for combination of decision procedures overcomes the complexity limitations and restrictions on the theories imposed by the standard Nelson-Oppen approach.

• Ph.D. Thesis 2012 An Adaptive Fast Multipole Method-Based PDE Solver in Three Dimensions Langston, Matthew Harper Abstract | PDF

  Title: An Adaptive Fast Multipole Method-Based PDE Solver in Three Dimensions

  Candidate: Langston, Matthew Harper

  Advisor(s): Zorin, Denis

  Abstract:

  Many problems in scientific computing require the accurate and fast solution to a variety of elliptic PDEs. These problems become increasingly dif.cult in three dimensions when forces become non-homogeneously distributed and geometries are complex.

  We present an adaptive fast volume solver using a new version of the fast multipole method, incorporated with a pre-existing boundary integral formulation for the development of an adaptive embedded boundary solver.

  For the fast volume solver portion of the algorithm, we present a kernel-independent, adaptive fast multipole method of arbitrary order accuracy for solving elliptic PDEs in three dimensions with radiation boundary conditions. The algorithm requires only a Green’s function evaluation routine for the governing equation and a representation of the source distribution (the right-hand side) that can be evaluated at arbi­trary points.

  The performance of the method is accelerated in two ways. First, we construct a piecewise polynomial approximation of the right-hand side and compute far-.eld expansions in the FMM from the coef.cients of this approximation. Second, we precompute tables of quadratures to handle the near-.eld interactions on adaptive octree data structures, keeping the total storage requirements in check through the exploitation of symmetries. We additionally show how we extend the free-space volume solver to solvers with periodic and well as Dirichlet boundary conditions.

  For incorporation with the boundary integral solver, we develop interpolation methods to maintain the accuracy of the volume solver. These methods use the existing FMM-based octree structure to locate ap­propriate interpolation points, building polynomial approximations to this larger set of forces and evaluating these polynomials to the locally under-re.ned grid in the area of interest.

  We present numerical examples for the Laplace, modi.ed Helmholtz and Stokes equations for a variety of boundary conditions and geometries as well as studies of the interpolation procedures and stability of far-.eld and polynomial constructions.

• Ph.D. Thesis 2012 Acquiring information from wider scope to improve event extraction Liao, Shasha Abstract | PDF

  Title: Acquiring information from wider scope to improve event extraction

  Candidate: Liao, Shasha

  Advisor(s): Grishman, Ralph

  Abstract:

  Event extraction is a particularly challenging type of information extraction (IE). Most current event extraction systems rely on local information at the phrase or sentence level. However, this local context may be insufficient to resolve ambiguities in identifying particular types of events; information from a wider scope can serve to resolve some of these ambiguities.

  In this thesis, we first investigate how to extract supervised and unsupervised features to improve a supervised baseline system. Then, we present two additional tasks to show the benefit of wider scope features in semi-supervised learning (self-training) and active learning (co-testing). Experiments show that using features from wider scope can not only aid a supervised local event extraction baseline system, but also help the semi-supervised or active learning approach.

• Ph.D. Thesis 2012 Mobile Accessibility Tools for the Visually Impaired Paisios, Nektarios Abstract | PDF

  Title: Mobile Accessibility Tools for the Visually Impaired

  Candidate: Paisios, Nektarios

  Advisor(s): Subramanian; Lakshminarayanan

  Abstract:

  Visually impaired users are in dire need of better accessibility tools. The past few years have witnessed an exponential growth in the computing capabilities and onboard sensing capabilities of mobile phones making them an ideal candidate for building next-generation applications. We believe that the mobile device can play a significant role in the future for aiding visually impaired users in day-to-day activities with simple and usable mobile accessibility tools. This thesis describes the design, implementation, evaluation and user-study based analysis of four different mobile accessibility applications.

  Our first system is the design of a highly accurate and usable mobile navigational guide that uses Wi-Fi and accelerometer sensors to navigate unfamiliar environments. A visually impaired user can use the system to construct a virtual topological map across points of interest within a building based on correlating the user' walking patterns (with turn signals) with the Wi-Fi and accelerometer readings. The user can subsequently use the map to navigate previously traveled routes. Our second system, Mobile Brailler, presents several prototype methods of text entry on a modern touch screen mobile phone that are based on the Braille alphabet and thus are convenient for visually impaired users. Our third system enables visually impaired users to leverage the camera of a mobile device to accurately recognize currency bills even if the images are partially or highly distorted. The final system enables visually impaired users to determine whether a pair of clothes, in this case of a tie and a shirt, can be worn together or not, based on the current social norms of color-matching.

  We believe that these applications together, provide a suite of important mobile accessibility tools to enhance four critical aspects of a day-to-day routine of a visually impaired user: to navigate easily, to type easily, to recognize currency bills (for payments) and to identify matching clothes.

• Ph.D. Thesis 2012 Reusable Software Infrastructure for Stream Processing Soule, Robert Abstract | PDF

  Title: Reusable Software Infrastructure for Stream Processing

  Candidate: Soule, Robert

  Advisor(s): Grimm, Robert

  Abstract:

  Developers increasingly use streaming languages to write their data processing applications. While a variety of streaming languages exist, each targeting a particular application domain, they are all similar in that they represent a program as a graph of streams (i.e. sequences of data items) and operators (i.e. data transformers). They are also similar in that they must process large volumes of data with high throughput. To meet this requirement, compilers of streaming languages must provide a variety of streaming-specific optimizations, including automatic parallelization. Traditionally, when many languages share a set of optimizations, language implementors translate the source languages into a common representation called an intermediate language (IL). Because optimizations can modify the IL directly, they can be re-used by all of the source languages, reducing the overall engineering effort. However, traditional ILs and their associated optimizations target single-machine, single-process programs. In contrast, the kinds of optimizations that compilers must perform in the streaming domain are quite different, and often involve reasoning across multiple machines. Consequently, existing ILs are not suited to streaming languages.

  This thesis addresses the problem of how to provide a reusable infrastructure for stream processing languages. Central to the approach is the design of an intermediate language specifically for streaming languages and optimizations. The hypothesis is that an intermediate language designed to meet the requirements of stream processing can assure implementation correctness; reduce overall implementation effort; and serve as a common substrate for critical optimizations. In evidence, this thesis provides the following contributions: (1) a catalog of common streaming optimizations that helps define the requirements of a streaming IL; (2) a calculus that enables reasoning about the correctness of source language translation and streaming optimizations; and (3) an intermediate language that preserves the semantics of the calculus, while addressing the implementation issues omitted from the calculus This work significantly reduces the effort it takes to develop stream processing languages, and jump-starts innovation in language and optimization design.

   

• Ph.D. Thesis 2012 Building scalable geo-replicated storage backends for web applications Sovran, Yair Abstract | PDF

  Title: Building scalable geo-replicated storage backends for web applications

  Candidate: Sovran, Yair

  Advisor(s): Li, Jinyang

  Abstract:

  Web applications increasingly require a storage system that is both scalable and can replicate data across many distant data centers or sites. Most existing storage solutions fall into one of two categories: Traditional databases offer strict consistency guarantees and programming ease, but are difficult to scale in a geo-replicated setting. NoSQL stores are scalable and efficient, but have weak consistency guarantees, placing the burden of ensuring consistency on programmers. In this dissertation, we describe two systems that help bridge the two extremes, providing scalable, geo-replicated storage for web applications, while also easy to program for. Walter is a key-value store that supports transactions and replicating data across distant sites. A key feature underlying Walter is a new isolation property: Parallel Snapshot Isolation (PSI). PSI allows Walter to replicate data asynchronously, while providing strong guarantees within each site. PSI does not allow write-write conflicts, alleviating the burden of writing conflict resolution logic. To prevent write-write conflicts and implement PSI, Walter uses two new and simple techniques: preferred sites and counting sets. Lynx is a distributed database backend for scaling latency-sensitive web applications. Lynx supports optimizing queries via data denormalization, distributed secondary indexes, and materialized join views. To preserve data constraints across denormalized tables and secondary indexes, Lynx relies on the a novel primitive: Distributed Transaction Chain (DTC). A DTC groups a sequence of transactions to be executed on different nodes while providing two guarantees. First, all transactions in a DTC execute exactly once despite failures. Second, transactions from concurrent DTCs are interleaved consistently on common nodes. We built several web applications on top of Walter and Lynx: an auction service, a microblogging service, and a social networking website. We have found that building web applications using Walter and Lynx is quick and easy. Our experiments show that the resulting applications are capable of providing scalable, low latency operation across multiple geo-replicated sites.

• Ph.D. Thesis 2012 Rapid Training of Information Extraction with Local and Global Data Views Sun, Ang Abstract | PDF

  Title: Rapid Training of Information Extraction with Local and Global Data Views

  Candidate: Sun, Ang

  Advisor(s): Grishman, Ralph

  Abstract:

  This dissertation focuses on fast system development for Information Extraction (IE). State-of-the-art systems heavily rely on extensively annotated corpora, which are slow to build for a new domain or task. Moreover, previous systems are mostly built with local evidence such as words in a short context window or features that are extracted at the sentence level. They usually generalize poorly on new domains.

  This dissertation presents novel approaches for rapidly training an IE system for a new domain or task based on both local and global evidence. Specifically, we present three systems: a relation type extension system based on active learning, a relation type extension system based on semi-supervised learning, and a cross-domain bootstrapping system for domain adaptive named entity extraction.

  The active learning procedure adopts features extracted at the sentence level as the local view and distributional similarities between relational phrases as the global view. It builds two classifiers based on these two views to find the most informative contention data points to request human labels so as to reduce annotation cost.

  The semi-supervised system aims to learn a large set of accurate patterns for extracting relations between names from only a few seed patterns. It estimates the confidence of a name pair both locally and globally: locally by looking at the patterns that connect the pair in isolation; globally by incorporating the evidence from the clusters of patterns that connect the pair. The use of pattern clusters can prevent semantic drift and contribute to a natural stopping criterion for semi-supervised relation pattern discovery.

  For adapting a named entity recognition system to a new domain, we propose a cross-domain bootstrapping algorithm, which iteratively learns a model for the new domain with labeled data from the original domain and unlabeled data from the new domain. We first use word clusters as global evidence to generalize features that are extracted from a local context window. We then select self-learned instances as additional training examples using multiple criteria, including some based on global evidence.

• Ph.D. Thesis 2012 Combating Sybil attacks in cooperative systems Tran, Nguyen Abstract | PDF

  Title: Combating Sybil attacks in cooperative systems

  Candidate: Tran, Nguyen

  Advisor(s): Li, Jinyang

  Abstract:

  Cooperative systems are ubiquitous nowadays. In a cooperative system, end users contribute resource to run the service instead of only receiving the service passively from the system. For example, users upload and comment pictures and videos on Flicker and YouTube, users submit and vote on news articles on Digg. As another example, users in BitTorrent contribute bandwidth and storage to help each other download content. As long as users behave as expected, these systems benefit immensely from user contribution. In fact, five out of ten most popular websites are operating in this cooperative fashion (Facebook, YouTube, Blogger, Twitter, Wikipedia). BitTorrent is dominating the global Internet traffic.

  A robust cooperative system cannot blindly trust that its users will truthfully participate in the system. Malicious users seek to exploit the systems for profit. Selfish users consume but avoid to contribute resource. For example, adversaries have manipulated the voting system of Digg to promote their articles of dubious quality. Selfish users in public BitTorrent communities leave the system to avoid uploading files to others, resulting in drastic performance degradation for these content distribution systems. The ultimate way to disrupt security and incentive mechanisms of cooperative systems is using Sybil attacks, in which the adversary creates many Sybil identities (fake identities) and use them to disrupt the systems' normal operation. No security and incentive mechanism works correctly if the systems do not have a robust identity management that can defend against Sybil attacks.

  This thesis provides robust identity management schemes which are resilient to the Sybil attack, and use them to secure and incentivize user contribution in several example cooperative systems. The main theme of this work is to leverage the social network among users in designing secure and incentive-compatible cooperative systems. First, we develop a distributed admission control protocol, called Gatekeeper, that leverages social network to admit most honest user identities and only few Sybil identities into the systems. Gatekeeper can be used as a robust identity management for both centralized and decentralized cooperative systems. Second, we provide a vote aggregation system for content voting systems, called SumUp, that can prevent an adversary from casting many bogus votes for a piece of content using the Sybil attack. SumUp leverages unique properties of content voting systems to provide significantly better Sybil defense compared with applying a general admission control protocol such as \gatekeeper. Finally, we provide a robust reputation system, called Credo, that can be used to incentivize bandwidth contribution in peer-to-peer content distribution networks. Credo reputation can capture user contribution, and is resilient to both Sybil and collusion attacks.

• Ph.D. Thesis 2012 Multi-species biclustering: An integrative method to identify functional gene conservation between multiple species Waltman, Peter Abstract | PDF

  Title: Multi-species biclustering: An integrative method to identify functional gene conservation between multiple species

  Candidate: Waltman, Peter

  Advisor(s): Bonneau, Richard

  Abstract:

  Background : Several recent comparative functional genomics projects have indicated that the co-regulation of many genes is conserved across species, at least in part. This suggests that comparative analysis of functional genomics data-sets could prove powerful in identifying co-regulated groups that are conserved across multiple species.

  Results : We present recent work to extend our cMonkey algorithm to simultaneously bicluster heterogeneous data from multiple species to identify conserved modules of orthologous genes, which can yield evolutionary insights into the formation of regulatory modules. We also present results from the multi-species analysis to two triplets of bacteria. The first of these is a triplet of Gram-positive bacteria consisting of Bacillus subtilis, Bacillus anthracis, and Listeria monocytogenes, while the second is a triplet of Gram-negative bacteria that includes Escherichia coli, Salmonella typhimurium and Vibrio cholerae. Finally, we will present initial results from the multi-species biclustering analysis of human and mouse hematopoietic differentiation data.

  Conclusion : Analysis of biclusters obtained revealed a surprising number of gene groups with conserved modularity and high biological significance as judged by several measures of cluster quality. We also highlight cases of interest from the Gram-positive triplet, including one that suggests a temporal difference in the expression of genes governing sporulation in the two Bacillus species. While analysis of the mouse and human hematopoietic differentiation is preliminary, it indicates the applicability of this analysis to eukaryotic systems, including comparison of cancer model systems. Finally, we suggest ways in which this analysis could be extended to identify divergent modules that may exist between normal and disease tissue.

• Ph.D. Thesis 2011 Collusion Preserving Computation Alwen, Joel Abstract | PDF

  Title: Collusion Preserving Computation

  Candidate: Alwen, Joel

  Advisor(s): Dodis, Yevgeniy

  Abstract:

  In collusion-free protocols, subliminal communication is impossible and parties are thus unable to communicate any information beyond what the protocol allows". Collusion-free protocols are interesting for several reasons, but have specifically attracted attention because they can be used to reduce trust in game-theoretic mechanisms. Collusion-free protocols are impossible to achieve (in general) when all parties are connected by point-to-point channels, but exist under certain physical assumptions (Lepinksi et al., STOC 2005) or in specific network topologies (Alwen et al., Crypto 2008).

  
  In addition to proposing the definition, we explore necessary properties of the underlying communication resource. Next we provide a general feasibility result for collusion-preserving computation of arbitrary functionalities. We show that the resulting protocols enjoy an elegant (and surprisingly strong) fallback security even in the case when the underlying communication resource acts in a Byzantine manner. Finally, we investigate the implications of these results in the context of mechanism design.

• Ph.D. Thesis 2011 Re-architecting Web and Mobile Information Access for Emerging Regions Chen, Jay Abstract | PDF

  Title: Re-architecting Web and Mobile Information Access for Emerging Regions

  Candidate: Chen, Jay

  Advisor(s): Subramanian; Lakshminarayanan

  Abstract:

  Providing access to information for people in emerging regions is an important problem. Over the past decade there have been many proposed and increasingly numerous deployed systems to enable information access, but successes are few and modest at best. Internet in emerging regions is still generally unusable or intolerably slow. Mobile phone applications are either not designed for the phones that poor people own, otherwise, the applications lack functionality, are difficult to use, or expensive to operate. In this work we focus on enabling digital information access for people in emerging regions.

  To advance the state of the art, we contribute numerous observations about how people access information in emerging regions, why the current models for web access and SMS platforms are broken, and techniques to enable applications over constrained Internet or SMS. The mechanisms presented here were designed after extensive field work in several different regions including rural, peri-urban, and urban areas in India, Kenya, Ghana, and Mexico. Multiple user studies were conducted throughout the course of system design and prototyping. We present a novel set of context appropriate platforms and tools, some spanning several layers of the networking stack. Five complete systems were implemented and deployed in the field. First, Event Logger for Firefox (ELF) is an easily deployable Firefox extension which functions as both a web browsing analysis tool and an in-browser web optimization platform. Second, RuralCafe provides a platform for web search and browsing over extremely slow or intermittent networks. Third, Contextual Information Portals (CIP) provide cached repositories of web pages tailored to the particular context in which it is to be used. Fourth, UjU is a mobile application platform that simplies the design of new SMS-based mobile applications. Finally, SMSFind is a SMS-based search service that runs on mobile phones without setup or subscription to a data plan.
  
  Taken as a whole, the systems here are a comprehensive solution for addressing the problem of enabling digital information access in emerging regions.

• Ph.D. Thesis 2011 Automatic Deduction for Theories of Algebraic Data Types Chikanian, Igor Abstract | PDF

  Title: Automatic Deduction for Theories of Algebraic Data Types

  Candidate: Chikanian, Igor

  Advisor(s): Barrett, Clark

  Abstract:

  In this thesis we present formal logical systems, concerned with reasoning about algebraic data types.

  The first formal system is based on the quantifier-free calculus (outermost universally quantified). This calculus is comprised of state change rules, and computations are performed by successive applications of these rules. Thereby, our calculus gives rise to an abstract decision procedure. This decision procedure determines if a given formula involving algebraic type members is valid. It is shown that this calculus is sound and complete. We also examine how this system performs practically and give experimental results. Our main contribution, as compared to previous work on this subject,is a new and more efficient decision procedure for checking satisfiability of the universal fragment within the theory of algebraic data types.

  The second formal system, called Term Builder, is the deductive system based on higher order type theory, which subsumes second order and higher order logics. The main purpose of this calculus is to formulate and prove theorems about algebraic or other arbitrary user-defined types.Term Builder supports proof objects and is both, an interactive theorem prover, and verifier. We describe the built-in deductive capabilities of Term Builder and show its consistency. The logic represented by our prover is intuitionistic. Naturally, it is also incomplete and undecidable, but its expressive power is much higher than that of the first formal system.

  Among our achievements in building this theorem prover is an elegant and intuitive GUI for building proofs. Also, a new feature from the foundational viewpoint is that, in contrast with other approaches, we have uniqueness-of-types property, which is not modulo beta-conversion.

• Ph.D. Thesis 2011 Efficient Cryptographic Primitives for Non-Interactive Zero-Knowledge Proofs and Applications Haralambiev, Kristiyan Abstract | PDF

  Title: Efficient Cryptographic Primitives for Non-Interactive Zero-Knowledge Proofs and Applications

  Candidate: Haralambiev, Kristiyan

  Advisor(s): Shoup, Victor

  Abstract:

  Non-interactive zero-knowledge (NIZK) proofs have enjoyed much interest in cryptography since they were introduced more than twenty years ago by Blum et al. [BFM88]. While quite useful when designing modular cryptographic schemes, until recently NIZK could be realized efficiently only using certain heuristics. However, such heuristic schemes have been widely criticized. In this work we focus on designing schemes which avoid them. In [GS08], Groth and Sahai presented the first efficient (and currently the only) NIZK proof system in the standard model. The construction is based on bilinear maps and is limited to languages of certain satisfiable system of equations. Given this expressibility limitation of the system of equations, we are interested in cryptographic primitives that are "compatible" with it. Equipped with such primitives and Groth-Sahai proof system, we show how to construct cryptographic schemes efficiently in a modular fashion. 

  In this work, we describe properties required by any cryptographic scheme to mesh well with Groth-Sahai proofs. Towards this, we introduce the notion of "structure-preserving" cryptographic scheme. We present the first constant-size structure-preserving signature scheme for messages consisting of general bilinear group elements. This allows us (for the first time) to instantiate efficiently a modular construction of round-optimal blind signature based on the framework of Fischlin [Fis06].
  
  Our structure-preserving homomorphic trapdoor commitment schemes yield efficient leakage-resilient signatures (in the bounded leakage model) which satisfy the standard security requirements and additionally tolerates any amount of leakage; all previous works satisfied at most two of those three properties.
  
  Lastly, we build a structure-preserving encryption scheme which satisfies the standard CCA security requirements. While somewhat similar to the notion of verifiable encryption, it provides better properties and yields the first efficient two-party protocol for joint ciphertext computation. Note that the efficient realization of such a protocol was not previously possible even using the heuristics mentioned above.
  
  Lastly, in this line of work, we revisit the notion of simulation extractability and define "true-simulation extractable" NIZK proofs. Although quite similar to the notion of simulation-sound extractable NIZK proofs, there is a subtle but rather important difference which makes it weaker and easier to instantiate efficiently. As it turns out, in many scenarios, this new notion is sufficient, and using it, we can construct efficient leakage resilient signatures and CCA encryption scheme.

   

• Ph.D. Thesis 2011 Learning Feature Hierarchies for Object Recognition Kavukcuoglu, Koray Abstract | PDF

  Title: Learning Feature Hierarchies for Object Recognition

  Candidate: Kavukcuoglu, Koray

  Advisor(s): LeCun, Yann

  Abstract:

  In this thesis we study unsupervised learning algorithms for training feature extractors and building deep learning models. We propose sparse-modeling algo- rithms as the foundation for unsupervised feature extraction systems. To reduce the cost of the inference process required to obtain the optimal sparse code, we model a feed-forward function that is trained to predict this optimal sparse code. Using an efficient predictor function enables the use of sparse coding in hierarchical models for object recognition. We demonstrate the performance of the developed system on several recognition tasks, including object recognition, handwritten digit classification and pedestrian detection. Robustness to noise or small variations in the input is a very desirable property for a feature extraction algorithm. In order to train locally-invariant feature extractors in an unsupervised manner, we use group sparsity criteria that promote similarity between the dictionary elements within a group. This model produces locally-invariant representations under small pertur- bations of the input, thus improving the robustness of the features. Many sparse modeling algorithms are trained on small image patches that are the same size as the dictionary elements. This forces the system to learn multiple shifted versions of each dictionary element. However, when used convolutionally over large im- ages to extract features, these models produce very redundant representations. To avoid this problem, we propose convolutional sparse coding algorithms that yield a richer set of dictionary elements, reduce the redundancy of the representation and improve recognition performance.

   

• Ph.D. Thesis 2011 Topics in Formal Synthesis and Modeling Klein, Uri Abstract | PDF

  Title: Topics in Formal Synthesis and Modeling

  Candidate: Klein, Uri

  Advisor(s): Pnueli, Amir; Zuck, Lenore

  Abstract:

  The work presented focuses on two problems, that of synthesizing systems from formal specifications, and that of formalizing REST -- a popular web applications' development pattern. 

  For the synthesis problem, we distinguish between the synchronous and the asynchronous case. For the former, we solve a problem concerning a fundamental flaw in specification construction in previous work. We continue with exploring effective synthesis of asynchronous systems (programs on multi-threaded systems). Two alternative models of asynchrony are presented, and shown to be equally expressive for the purpose of synthesis.
  
  REST is a software architectural style used for the design of highly scalable web applications. Interest in REST has grown rapidly over the past decade. However, there is also considerable confusion surrounding REST: many examples of supposedly RESTful APIs violate key REST constraints. We show that the constraints of REST and of RESTful HTTP can be precisely formulated within temporal logic. This leads to methods for model checking and run-time verification of RESTful behavior. We formulate several relevant verification questions and analyze their complexity. 

• Ph.D. Thesis 2011 Adaptive Isotopic Approximation of Nonsingular Curves and Surfaces Lin, Long Abstract | PDF

  Title: Adaptive Isotopic Approximation of Nonsingular Curves and Surfaces

  Candidate: Lin, Long

  Advisor(s): Yap, Chee

  Abstract:

  Consider the problem of computing isotopic approximations of nonsingular curves and surfaces that are implicitly represented by equations of the form f (X, Y )=0 and f (X,Y, Z)=0. Thisfundamentalproblem has seen much progress along several fronts, but we will focus on domain subdivision algorithms. Two algorithms in this area are from Snyder(1992) and Plantinga and Vegter(2004). We introduce a family of new algorithms that combines the advantages of these two algorithms: like Snyder, we use the parameterizability criterion for subdivision, and like Plantinga and Vegter, we exploit nonlocal isotopy.

  We first apply our approach to curves, resulting in a more efficient algorithm. We then extend our approach to surfaces. The extension is by no means routine, as the correctness arguments and case analysis are more subtle. Also, a new phenomenon arises in which local rules for constructing surfaces are no longer sufficient.

  We further extend our algorithms in two important and practical directions: first, we allow subdivision cells to be non squares or non cubes, with arbitrary but bounded aspect ratios: in 2D, we allow boxes to be split into 2 or 4 children; and in 3D, we allow boxes to be split into 2, 4 or 8 children. Second, we allow the inputregion-of-interest(ROI) to have arbitrary geometry represented by anquadtreeoroctree,aslongas the curves or surfaces has no singularities in the ROI and intersects the boundary of ROI transversally.

  Our algorithm is numerical because our primitives are based on interval arithmetic and exact BigFloat numbers. It is practical, easy to implement exactly (compared to algebraic approaches) and does not suffer from implementation gaps (compared to geometric approaches). We report some very encouraging experimental results,showing that our algorithms can be much more efficient than the algorithms of Plantinga and Vegter(2D and 3D)and Snyder(2D only).

• Ph.D. Thesis 2011 Real-Space Localization Methods for Minimizing the Kohn-Sham Energy Millstone, Marc Abstract | PDF

  Title: Real-Space Localization Methods for Minimizing the Kohn-Sham Energy

  Candidate: Millstone, Marc

  Advisor(s): Overton, Michael

  Abstract:

  The combination of ever increasing computational power and new mathematical models has fundamentally changed the field of computational chemistry. One example of this is the use of new algorithms for computing the charge density of a molecular system from which one can predict many physical properties of the system.

  This thesis presents two new algorithms for minimizing the Kohn-Sham energy, which is used to describe a system of non-interacting electrons through a set of single-particle wavefunctions. By exploiting a known localization region of the wavefunctions, each algorithm evaluates the Kohn-Sham energy function and gradient at a set of iterates that have a special sparsity structure. We have chosen to represent the problem in real-space using finite-differences, allowing us to efficiently evaluate the energy function and gradient using sparse linear algebra. Detailed numerical experiments are provided on a set of representative molecules demonstrating the performance and robustness of these methods.

• Ph.D. Thesis 2011 Scoring-and-Unfolding Trimmed Tree Assembler: Algorithms for Assembling Genome Sequences Accurately and Efficiently Narzisi, Giuseppe Abstract | PDF

  Title: Scoring-and-Unfolding Trimmed Tree Assembler: Algorithms for Assembling Genome Sequences Accurately and Efficiently

  Candidate: Narzisi, Giuseppe

  Advisor(s): Mishra, Bud

  Abstract:

  The recent advances in DNA sequencing technology and their many potential applications to Biology and Medicine have rekindled enormous interest in several classical algorithmic problems at the core of Genomics and Computational Biology: primarily, the whole-genome sequence assembly problem (WGSA). Two decades back, in the context of the Human Genome Project, the problem had received unprecedented scientific prominence: its computational complexity and intractability were thought to have been well understood; various competitive heuristics, thoroughly explored and the necessary software, properly implemented and validated. However, several recent studies, focusing on the experimental validation of de novo assemblies, have highlighted several limitations of the current assemblers.

  Intrigued by these negative results, this dissertation reinvestigates the algorithmic techniques required to correctly and efficiently assemble genomes. Mired by its connection to a well-known NP-complete combinatorial optimization problem, historically, WGSA has been assumed to be amenable only to greedy and heuristic methods. By placing efficiency as their priority, these methods opted to rely on local searches, and are thus inherently approximate, ambiguous or error-prone. This dissertation presents a novel sequence assembler, SUTTA, that dispenses with the idea of limiting the solutions to just the approximated ones, and instead favors an approach that could potentially lead to an exhaustive (exponential-time) search of all possible layouts but tames the complexity through constrained search (Branch-and-Bound) and quick identification and pruning of implausible solutions.

  Complementary to this problem is the task of validating the generated assemblies. Unfortunately, no commonly accepted method exists yet and widely used metrics to compare the assembled sequences emphasize only size, poorly capturing quality and accuracy. This dissertation also addresses these concerns by developing a more comprehensive metric, the Feature-Response Curve, that, using ideas from classical ROC (receiver-operating characteristic) curve, more faithfully captures the trade-off between contiguity and quality.

  Finally, this dissertation demonstrates the advantages of a complete pipeline integrating base-calling (TotalReCaller) with assembly (SUTTA) in a Bayesian manner.

• Ph.D. Thesis 2011 Cryptographic Resilience to Continual Information Leakage Wichs, Daniel Abstract | PDF

  Title: Cryptographic Resilience to Continual Information Leakage

  Candidate: Wichs, Daniel

  Advisor(s): Dodis, Yevgeniy

  Abstract:

  We study the question of achieving cryptographic security on devices that leak information about their internal secret state to an external attacker.This study is motivated by the prevalence of side-channel attacks, where the physical characteristics of a computation (e.g. timing, power-consumption, temperature, radiation, acoustics, etc.) can be measured, and may reveal useful information about the internal state of a device. Since some such leakage is inevitably present in almost any physical implementation, we believe that this problem cannot just be addressed by physical countermeasures alone. Instead, it should already be taken into account when designing the mathematical specification of cryptographic primitives and included in the formal study of their security.

  In this thesis, we propose a new formal framework for modeling the leakage available to an attacker. This framework, called the continual leakage model, assumes that an attacker can continually learn arbitrary information about the internal secret state of a cryptographic scheme at any point in time, subject only to the constraint that the rate of leakage is bounded. More precisely, our model assumes some abstract notion of time periods. In each such period, the attacker can choose to learn arbitrary functions of the current secret state of the scheme, as long as the number of output bits leaked is not too large. In our solutions, cryptographic schemes will continually update their internal secret state at the end of each time period. This will ensure that leakage observed in different time periods cannot be meaningfully combined to break the security of the cryptosystem. Although these updates modify the secret state of the cryptosystem, the desired functionality of the scheme is preserved, and the users can remain oblivious to these updates. We construct signatures, encryption, and secret sharing/storage schemes in this model.

• Ph.D. Thesis 2011 Surface Representation of Particle Based Fluids Yu, Jihun Abstract | PDF

  Title: Surface Representation of Particle Based Fluids

  Candidate: Yu, Jihun

  Advisor(s): Yap, Chee

  Abstract:

  In this thesis, we focus on surface representation for particle-based fluid simulators such as Smoothed Particle Hydrodynamics (SPH). We first present a new surface reconstruction algorithm which formulates the implicit function as a sum of anisotropic smoothing kernels. The direction of anisotropy at a particle is determined by performing Weighted Principal Component Analysis (WPCA) over the neighboring particles. In addition, we perform a smoothing step that re-positions the centers of these smoothing kernels. Since these anisotropic moothing kernels capture the local particle distributions more accurately, our method has advantages over existing methods in representing smooth surfaces, thin streams and sharp features of fluids. This method is fast, easy to implement, and the results demonstrate a significant improvement in the quality of reconstructed surfaces as compared to existing methods. Next,we introduce the idea of using an explicit triangle mesh to track the air/liquid interface in a SPH simulator.

  Once an initial surface mesh is created, this mesh is carried forward in time using nearby particle velocities to advect the mesh vertices. The mesh connectivity remains mostly unchanged across time-steps; it is only modified locally for topology change events or for the improvement of triangle quality. In order to ensure that the surface mesh does not diverge from the underlying particle simulation, we periodically project the mesh surface onto an implicit surface defined by the physics simulation. The mesh surface presents several advantages over previous SPH surface tracking techniques: A new method for surface tension calculations clearly outperforms the state of the art in SPH surface tension for computer graphics. A new method for tracking detailed surface information (like colors) is less susceptible to numerical diffusion than competing techniques. Finally, a temporally-coherent surface mesh allows us to simulate high-resolution surface wave dynamics without being limited by the particle resolution of the SPH simulation.

• Ph.D. Thesis 2010 On the Randomness Requirements for Privacy Bosley, Carleton Abstract | PDF

  Title: On the Randomness Requirements for Privacy

  Candidate: Bosley, Carleton

  Advisor(s): Dodis, Yevgeniy

  Abstract:

  Most cryptographic primitives require randomness (for example, to generate secret keys). Usually, one assumes that perfect randomness is available, but, conceivably, such primitives might be built under weaker, more realistic assumptions. This is known to be achievable for many authentication applications, when entropy alone is typically sufficient. In contrast, all known techniques for achieving privacy seem to fundamentally require (nearly) perfect randomness. We ask the question whether this is just a coincidence, or, perhaps, privacy inherently requires true randomness?

  We completely resolve this question for information-theoretic private-key encryption, where parties wish to encrypt a b-bit value using a shared secret key sampled from some imperfect source of randomness S. Our technique also extends to related primitives which are sufficiently binding and hiding, including computationally secure commitments and public-key encryption.

  Our main result shows that if such n-bit source S allows for a secure encryption of b bits, where b > log n, then one can deterministically extract nearly b almost perfect random bits from S . Further, the restriction that b > log n is nearly tight: there exist sources S allowing one to perfectly encrypt (log n - log log n) bits, but not to deterministically extract even a single slightly unbiased bit.

  Hence, to a large extent, true randomness is inherent for encryption: either the key length must be exponential in the message length b, or one can deterministically extract nearly b almost unbiased random bits from the key. In particular, the one-time pad scheme is essentially "universal".

• Ph.D. Thesis 2010 Machine Learning Approaches to Gene Duplication and Transcription Regulation Chen, Huang-Wen Abstract | PDF

  Title: Machine Learning Approaches to Gene Duplication and Transcription Regulation

  Candidate: Chen, Huang-Wen

  Advisor(s): Shasha, Dennis

  Abstract:

  Gene duplication can lead to genetic redundancy or functional divergence, when duplicated genes evolve independently or partition the original function. In this dissertation, we employed machine learning approaches to study two different views of this problem: 1) Redundome, which explored the redundancy of gene pairs in the genome of Arabidopsis thaliana, and 2) ContactBind, which focused on functional divergence of transcription factors by mutating contact residues to change binding affinity.

  In the Redundome project, we used machine learning techniques to classify gene family members into redundant and non-redundant gene pairs in Arabidopsis thaliana, where sufficient genetic and genomic data is available. We showed that Support Vector Machines were two-fold more precise than single attribute classifiers, and performed among the best within other machine learning algorithms. Machine learning methods predict that about half of all genes in Arabidopsis showed the signature of predicted redundancy with at least one but typically less than three other family members. Interestingly, a large proportion of predicted redundant gene pairs were relatively old duplications (e.g., Ks>1), suggesting that redundancy is stable over long evolutionary periods. The genome-wide predictions were plot with similarity trees based on ClustalW alignment scores, and can be accessed at http://redundome.bio.nyu.edu .

  In the ContactBind project, we use Bayesian networks to model dependences between contact residues in transcription factors and binding site sequences. Based on the models learned from various binding experiments, we predicted binding motifs and their locations on promoters for three families of transcription factors in three species. The predictions are publicly available at http://contactbind.bio.nyu.edu . The website also provides tools to predict binding motifs and their locations for novel protein sequences of transcription factors. Users can construct their Bayesian networks for new families once such a familial binding data is available.

• Ph.D. Thesis 2010 New Privacy-Preserving Architectures for Identity-/Attribute-based Encryption Chow, Sze Ming Abstract | PDF

  Title: New Privacy-Preserving Architectures for Identity-/Attribute-based Encryption

  Candidate: Chow, Sze Ming

  Advisor(s): Dodis, Yevgeniy; Shoup, Victor

  Abstract:

  The notion of identity-based encryption (IBE) was proposed as an economical alternative to public-key infrastructures. IBE is also a useful building block in various cryptographic primitives such as searchable encryption. A generalization of IBE is attribute-based encryption (ABE). A major application of ABE is fine-grained cryptographic access control of data. Research on these topics is still actively continuing.

  However, security and privacy of IBE and ABE are hinged on the assumption that the authority which setups the system is honest. Our study aims to reduce this trust assumption.

  The inherent key escrow of IBE has sparkled numerous debates in the cryptography/security community. A curious key generation center (KGC) can simply generate the user's private key to decrypt a ciphertext. However, can a KGC still decrypt if it does not know the intended recipient of the ciphertext? This question is answered by formalizing KGC anonymous ciphertext indistinguishability (ACI-KGC). All existing practical pairing-based IBE schemes without random oracles do not achieve this notion. In this thesis, we propose an IBE scheme with ACI-KGC, and a new system architecture with an anonymous secret key generation protocol such that the KGC can issue keys to authenticated users without knowing the list of users' identities. This also matches the practice that authentication should be done with the local registration authorities. Our proposal can be viewed as mitigating the key escrow problem in a new dimension.

  For ABE, it is not realistic to trust a single authority to monitor all attributes and hence distributing control over many attribute-authorities is desirable. A multi-authority ABE scheme can be realized with a trusted central authority (CA) which issues part of the decryption key according to a user's global identifier (GID). However, this CA may have the power to decrypt every ciphertext, and the use of a consistent GID allowed the attribute-authorities to collectively build a full profile with all of a user's attributes. This thesis proposes a solution without the trusted CA and without compromising users' privacy, thus making ABE more usable in practice.

  Underlying both contributions are our new privacy-preserving architectures enabled by borrowing techniques from anonymous credential.

• Ph.D. Thesis 2010 Tools and Techniques for the Sound Verification of Low Level Code Conway, Christopher L. Abstract | PDF

  Title: Tools and Techniques for the Sound Verification of Low Level Code

  Candidate: Conway, Christopher L.

  Advisor(s): Barrett, Clark

  Abstract:

  Software plays an increasingly crucial role in nearly every facet of modern life, from communications infrastructure to the control systems in automobiles, airplanes, and power plants. To achieve the highest degree of reliability for the most critical pieces of software, it is necessary to move beyond ad hoc testing and review processes towards verification---to prove using formal methods that a piece of code exhibits exactly those behaviors allowed by its specification and no others.

  A significant portion of the existing software infrastructure is written in low-level languages like C and C++. Features of these language present significant verification challenges. For example, unrestricted pointer manipulation means that we cannot prove even the simplest properties of programs without first collecting precise information about potential aliasing relationships between variables.

  In this thesis, I present several contributions. The first is a general framework for combining program analyses that are only conditionally sound. Using this framework, I show it is possible to design a sound verification tool that relies on a separate, previously-computed pointer analysis.

  The second contribution of this thesis is Cascade, a multi-platform, multi-paradigm framework for verification. Cascade includes a support for precise analysis of low-level C code, as well as for higher-level languages such as SPL.

  Finally, I describe a novel technique for the verification of datatype invariants in low-level systems code. The programmer provides a high-level specification for a low-level implementation in the form of inductive datatype declarations and code assertions. The connection between the high-level semantics and the implementation code is then checked using bit-precise reasoning. An implementation of this datatype verification technique is available as a Cascade module.

• Ph.D. Thesis 2010 Probabilistic and Topological methods in Computational Geometry Dhandapani, Raghavan Abstract | PDF

  Title: Probabilistic and Topological methods in Computational Geometry

  Candidate: Dhandapani, Raghavan

  Advisor(s): Pach, Janos

  Abstract:

  We consider four problems connected by the common thread of geometry. The first three involve problems and algorithms that arise in applications that apriori do not involve geometry but this turns out to be the right language for visualizing and analyzing them. In the fourth, we generalize some well known results in geometry to the topological plane. The techniques we use come from probability and topology.

  First, we consider two algorithms that work well in practice but the theoretical mechanism behind whose success is not very well understood.

  Greedy routing is a routing mechanism that is commonly used in wireless sensor networks. While routing on the Internet uses standard established protocols, routing in ad-hoc networks with little structure (like sensor networks) is more difficult. Practitioners have devised algorithms that work well in practice, however they were no known theoretical guarantees. We provide the first such result in this area by showing that greedy routing can be made to work on Planar triangulations.

  Linear Programming is a technique for optimizing a linear function subject to linear constraints. Simplex Algorithms are a family of algorithms that have proven quite successful in solving Linear Programs in practice. However, examples of Linear Programs on which these algorithms are very inefficient have been obtained by researchers. In order to explain this discrepancy between theory and practice, many authors have shown that Simplex Algorithms are efficient in expectation on randomized Linear Programs. We strengthen these results by proving a partial concentration bound for the Shadow Vertex Simplex Algorithm.

  Next, we point out a limitation in an algorithm that is used commonly by practitioners and suggest a way of overcoming this.

  Recommendation Systems are algorithms that are used to recommend goods (books, movies etc.) to users based on the similarities between their past preferences and those of other users. Low Rank Approximation is a common method used for this. We point out a common limitation of this method in the presence of ill-conditioning: the presence of multiple local minima. We also suggest a simple averaging based technique to overcome this limitation.

  Finally, we consider some basic results in convexity like Radon's, Helly's and Caratheodory's theorems and generalize them to the topological plane, i.e., a plane which has the concept of a linear path which is analogous to a straight line but no notion of metric or distances.

• Ph.D. Thesis 2010 Semi-Supervised Learning via Generalized Maximum Entropy Erkan, Ayse Naz Abstract | PDF

  Title: Semi-Supervised Learning via Generalized Maximum Entropy

  Candidate: Erkan, Ayse Naz

  Advisor(s): LeCun, Yann

  Abstract:

  Maximum entropy (MaxEnt) framework has been studied extensively in the supervised setting. Here, the goal is to find a distribution p, that maximizes an entropy function while enforcing data constraints so that the expected values of some (pre-defined) features with respect to p, match their empirical counterparts approximately. Using different entropy measures, different model spaces for p and different approximation criteria for the data constraints yields a family of discriminative supervised learning methods (e.g., logistic regression, conditional random fields, least squares and boosting). This framework is known as the generalized maximum entropy framework.

  Semi-supervised learning (SSL) has emerged in the last decade as a promising field that enables utilizing unlabeled data along with labeled data so as to increase the accuracy and robustness of inference algorithms. However, most SSL algorithms to date have had trade-offs, for instance in terms of scalability or applicability to multi-categorical data.

  In this thesis, we extend the generalized MaxEnt framework to develop a family of novel SSL algorithms using two different approaches: i. Introducing Similarity Constraints We incorporate unlabeled data via modifications to the primal MaxEnt objective in terms of additional potential functions. A potential function stands for a closed proper convex function that can take the form of a constraint and/or a penalty representing our structural assumptions on the data geometry. Specifically, we impose similarity constraints as additional penalties based on the semi-supervised smoothness assumption; i.e., we restrict the generalized MaxEnt problem such that similar samples have similar model outputs. ii. Augmenting Constraints on Model Features We incorporate unlabeled data to enhance the estimates on the model and empirical expectations based on our assumptions on the data geometry.

  In particular, we derive the semi-supervised formulations for three specific instances of the generalized MaxEnt on conditional distributions, namely logistic regression and kernel logistic regression for multi-class problems, and conditional random fields for structured output prediction problems. A thorough empirical evaluation on standard data sets that are widely used in the literature demonstrates the validity and competitiveness of the proposed algorithms. In addition to these benchmark data sets, we apply our approach to two real-life problems: i. vision based robot grasping, and ii. remote sensing image classification, where the scarcity of the labeled training samples is the main bottleneck in the learning process. For the particular case of grasp learning, we propose a combination of semi-supervised learning and active learning, another sub-field of machine learning that is focused on the scarcity of labeled samples, when the problem setup is suitable for incremental labeling.

  The novel SSL algorithms proposed in this thesis have numerous advantages over the existing semi-supervised algorithms as they yield convex, scalable, inherently multi-class loss functions that can be kernelized naturally.

• Ph.D. Thesis 2010 Solving Quantified First Order Formulas in Satisfiability Modulo Theories Ge, Yeting Abstract | PDF

  Title: Solving Quantified First Order Formulas in Satisfiability Modulo Theories

  Candidate: Ge, Yeting

  Advisor(s): Barrett, Clark

  Abstract:

  Design errors in computer systems, i.e. bugs, can cause inconvenience, loss of data and time, and in some cases catastrophic damages. One approach for improving design correctness is formal methods: techniques aiming at mathematically establishing that a piece of hardware or software satisfies certain properties. For some industrial cases in which formal methods are utilized, quantified first order formulas in satisfiability modulo theories (SMT) are useful. This dissertation presents several novel techniques for solving quantified formulas in SMT.

  In general, deciding a quantified formula in SMT is undecidable. The practical approach for general quantifier reasoning in SMT is heuristics-based instantiation. This dissertation proposes a number of new heuristics that solves several challenges. Experimental results show that with the new heuristics a significant number of more benchmarks can be solved than before.

  When only consider formulas within certain fragments of first order logic, it is possible to have complete algorithms based on instantiation. We propose several new fragments, and we prove that formulas in these fragments can be solved by a complete algorithm based on instantiation. For satisfiable quantified formulas in these fragments, we show how to construct the models.

  As SMT solvers grow in complexity, the correctness of SMT solvers become questionable. A practical method to improve the correctness is to check the proofs from SMT solvers. We propose a proof translator that translates proofs from SMT solver CVC3 into a trusted solver HOL Light that actually checks the proofs. Experiments with the proof translator discover a faulty proof rule in CVC3 and two MIT-labeled quantified benchmarks in the SMT benchmark library SMT-LIB.

• Ph.D. Thesis 2010 An Algorithmic Enquiry Concerning Causality Kleinberg, Samantha Abstract | PDF

  Title: An Algorithmic Enquiry Concerning Causality

  Candidate: Kleinberg, Samantha

  Advisor(s): Mishra, Bhubaneswar

  Abstract:

  In many domains we face the problem of determining the underlying causal structure from time-course observations of a system. Whether we have neural spike trains in neuroscience, gene expression levels in systems biology, or stock price movements in finance, we want to determine why these systems behave the way they do. For this purpose we must assess which of the myriad possible causes are significant while aiming to do so with a feasible computational complexity. At the same time, there has been much work in philosophy on what it means for something to be a cause, but comparatively little attention has been paid to how we can identify these causes. Algorithmic approaches from computer science have provided the first steps in this direction, but fail to capture the complex, probabilistic and temporal nature of the relationships we seek.

  This dissertation presents a novel approach to the inference of general (type-level) and singular (token-level) causes. The approach combines philosophical notions of causality with algorithmic approaches built on model checking and statistical techniques for false discovery rate control. By using a probabilistic computation tree logic to describe both cause and effect, we allow for complex relationships and explicit description of the time between cause and effect as well as the probability of this relationship being observed (e.g. "a and b until c, causing d in 10-20 time units"). Using these causal formulas and their associated probabilities, we develop a novel measure for the significance of a cause for its effect, thus allowing discovery of those that are statistically interesting, determined using the concepts of multiple hypothesis testing and false discovery control. We develop algorithms for testing these properties in time-series observations and for relating the inferred general relationships to token-level events (described as sequences of observations). Finally, we illustrate these ideas with example data from both neuroscience and finance, comparing the results to those found with other inference methods. The results demonstrate that our approach achieves superior control of false discovery rates, due to its ability to appropriately represent and infer temporal information.

• Ph.D. Thesis 2010 Time Series Modeling with Hidden Variables and Gradient-Based Algorithms Mirowski, Piotr Abstract | PDF

  Title: Time Series Modeling with Hidden Variables and Gradient-Based Algorithms

  Candidate: Mirowski, Piotr

  Advisor(s): LeCun, Yann

  Abstract:

  We collect time series from real-world phenomena, such as gene interactions in biology or word frequencies in consecutive news articles. However, these data present us with an incomplete picture, as they result from complex dynamical processes involving unobserved state variables. Research on state-space models is motivated by simultaneously trying to infer hidden state variables from observations, as well as learning the associated dynamic and generative models.

  I have developed a tractable, gradient-based method for training Dynamic Factor Graphs (DFG) with continuous latent variables. A DFG consists of (potentially nonlinear) factors modeling joint probabilities between hidden and observed variables. The DFG assigns a scalar energy to each configuration of variables, and a gradient-based inference procedure finds the minimum-energy state sequence for a given observation sequence. We approximate maximum likelihood learning by minimizing the expected energy over training sequences with respect to the factors' parameters. These alternated inference and parameter updates constitute a deterministic EM-like procedure.

  Using nonlinear factors such as deep, convolutional networks, DFGs were shown to reconstruct chaotic attractors, to outperform a time series prediction benchmark, and to successfully impute motion capture data where a large number of markers were missing. In a joint work with the NYU Plant Systems Biology Lab, DFGs have been subsequently employed to the discovery of gene regulation networks by learning the dynamics of mRNA expression levels.

  DFGs have also been extended into a deep auto-encoder architecture, and used on time-stamped text documents, with word frequencies as inputs. We focused on collections of documents that exhibit a structure over time. Working as dynamic topic models, DFGs could extract a latent trajectory from consecutive political speeches; applied to news articles, they achieved state-of-the-art text categorization and retrieval performance.

  Finally, I used an embodiment of DFGs to evaluate the likelihood of discrete sequences of words in text corpora, relying on dynamics on word embeddings. Collaborating with AT&T; Labs Research on a project in speech recognition, we have improved on existing continuous statistical language models by enriching them with word features and long-range topic dependencies.

• Ph.D. Thesis 2010 Structure Prediction and Visualization in Molecular Biology Poultney, Christopher Abstract | PDF

  Title: Structure Prediction and Visualization in Molecular Biology

  Candidate: Poultney, Christopher

  Advisor(s): Shasha, Dennis

  Abstract:

  The tools of computer science can be a tremendous help to the working biologist. Two broad areas where this is particularly true are visualization and prediction. In visualization, the size of the data involved often makes meaningful exploration of the data and discovery of salient features difficult and time-consuming. Similarly, intelligent prediction algorithms can greatly reduce the lab time required to achieve significant results, or can reduce an intractable space of potential experiments to a tractable size.

  Whereas the thesis discusses both a visualization technique and a machine learning problem, the thesis presentation will focus exclusively on the machine learning problem: prediction of temperature-sensitive mutations from protein structure. Temperature-sensitive mutations are a tremendously valuable research tool particularly for studying genes such as yeast essentially genes. To date, most methods for generating temperature-sensitive mutations involve large-scale random mutations followed by an intensive screening and characterization process. While there have been successful efforts to improve this process by rational design of temperature-sensitive proteins, surprisingly little work has been done in the area of predicting those mutations that will exhibit a temperature-sensitive phenotype. We describe a system that, given the structure of a protein of interest, uses a combination of protein structure prediction and machine learning to provide a ranked "top 5" list of likely candidates for temperature-sensitive mutations.

• Ph.D. Thesis 2010 Theoretical Foundations and Algorithms for Learning with Multiple Kernels Rostamizadeh, Afshin Abstract | PDF

  Title: Theoretical Foundations and Algorithms for Learning with Multiple Kernels

  Candidate: Rostamizadeh, Afshin

  Advisor(s): Mohri, Mehryar

  Abstract:

  Kernel-based algorithms have been used with great success in a variety of machine learning applications. These include algorithms such as support vector machines for classification, kernel ridge regression, ranking algorithms, clustering algorithms, and virtually all popular dimensionality reduction algorithms, since they are special instances of kernel principal component analysis.

  But, the choice of the kernel, which is crucial to the success of these algorithms, has been traditionally left entirely to the user. Rather than requesting the user to commit to a specific kernel, multiple kernel algorithms require the user only to specify a family of kernels. This family of kernels can be used by a learning algorithm to form a combined kernel and derive an accurate predictor. This is a problem that has attracted a lot of attention recently, both from the theoretical point of view and from the algorithmic, optimization, and application point of view.

  This thesis presents a number of novel theoretical and algorithmic results for learning with multiple kernels.

  It gives the first tight margin-based generalization bounds for learning kernels with Lp regularization. In particular, our margin bounds for L1 regularization are shown to have only a logarithmic dependency on the number of kernels, which is a significant improvement over all previous analyses. Our results also include stability-based guarantees for a class of regression algorithms. In all cases, these guarantees indicate the benefits of learning with a large number of kernels.

  We also present a family of new two-stage algorithms for learning kernels based on a notion of alignment and give an extensive analysis of the properties of these algorithms. We show the existence of good predictors for the notion of alignment we define and give efficient algorithms for learning a maximum alignment kernel by showing that the problem can be reduced to a simple QP.

  Finally, we also report the results of extensive experiments with our two-stage algorithms in classification and regression tasks, which show an improvement both over the uniform combination of kernels and over other state-of-the-art learning kernel methods for L1 and L2 regularization. These might constitute the first series of results for learning with multiple kernels that demonstrate a consistent improvement over a uniform combination of kernels.

• Ph.D. Thesis 2010 Creating collections and evaluating viewpoints: Selection techniques for interface design Secord, Adrian Abstract | PDF

  Title: Creating collections and evaluating viewpoints: Selection techniques for interface design

  Candidate: Secord, Adrian

  Advisor(s): Zorin, Denis

  Abstract:

  In computer graphics and user interface design, selection problems are those that require the user to select a collection consisting of a small number of items from a much larger library. This dissertation explores selection problems in two diverse domains: large personal multimedia collections, containing items such as personal photographs or songs, and camera positions for 3D objects, where each item is a different viewpoint observing an object. Multimedia collections have by discrete items with strong associated metadata, while camera positions form a continuous space but are weak in metadata. In either domain, the items to be selected have rich interconnections and dependencies, making it difficult to successfully apply simple techniques (such as ranking) to aid the user. Accordingly, we develop separate approaches for the two domains.

  For personal multimedia collections, we leverage the semantic metadata associated with each item (such as song title, artist name, etc.) and provide the user with a simple query language to describe their desired collection. Our system automatically suggests a collection of items that conform to the user’s query. Since any query language has limited expressive power, and since users often create collections via exploration, we provide various refinement techniques that allow the user to expand, refine and explore their collection directly through examples.

  For camera positioning, we do not have the advantage of having semantic metadata for each item, unlike in media collections. We instead create a proxy viewpoint goodness function which can be used to guide the solution of various selection problems involving camera viewpoints. This function is constructed from several different attributes of the viewpoint, such as how much surface area is visible, or how "curvy" the silhouette is. Since there are many possible viewpoint goodness functions, we conducted a large user study of viewpoint preference and use the results to evaluate thousands of different functions and find the best ones. While we suggest several goodness functions to the practitioner, our user study data and methodology can be used to evaluate any proposed goodness function; we hope it will be a useful tool for other researchers.

• Ph.D. Thesis 2010 Analysis of Mass Spectrometry Data for Protein Identification In Complex Biological Mixtures Spivak, Marina Abstract | PDF

  Title: Analysis of Mass Spectrometry Data for Protein Identification In Complex Biological Mixtures

  Candidate: Spivak, Marina

  Advisor(s): Greengard, Leslie

  Abstract:

  Mass spectrometry is a powerful technique in analytical chemistry that was originally designed to determine the composition of small molecules in terms of their constituent elements. In the last several decades, it has begun to be used for much more complex tasks, including the detailed analysis of the amino acid sequence that makes up an unknown protein and even the identification of multiple proteins present in a complex mixture. The latter problem is largely unsolved and the principal subject of this dissertation.

  The fundamental difficulty in the analysis of mass spectrometry data is that of ill-posedness. There are multiple solutions consistent with the experimental data and the data is subject to significant amounts of noise. In this work, we have developed application-specific machine learning algorithms that (partially) overcome this ill-posedness. We make use of labeled examples of a single class of peptide fragments and of the unlabeled fragments detected by the instrument. This places the approach within the broader framework of semi-supervised learning.

  Recently, there has been considerable interest in classification problems of this type, where the learning algorithm only has access to labeled examples of a single class and unlabeled data. The motivation for such problems is that in many applications, examples of one of the two classes are easy and inexpensive to obtain, whereas the acquisition of examples of a second class is difficult and labor-intensive. For example, in document classification, positive examples are documents that address specific subject, while unlabeled documents are abundant. In movie rating, the positive data are the movies chosen by clients, while the unlabeled data are all remaining movies in a collection. In medical imaging, positive (labeled) data correspond to images of tissue affected by a disease, while the remaining available images of the same tissue comprise the unlabeled data. Protein identification using mass spectrometry is another variant of such a general problem.

  In this work, we propose application-specific machine learning algorithms to address this problem. The reliable identification of proteins from mixtures using mass spectrometry would provide an important tool in both biomedical research and clinical practice.

• Ph.D. Thesis 2010 Matrix Approximation for Large-scale Learning Talwalkar, Ameet Abstract | PDF

  Title: Matrix Approximation for Large-scale Learning

  Candidate: Talwalkar, Ameet

  Advisor(s): Mohri, Mehryar

  Abstract:

  Modern learning problems in computer vision, natural language processing, computational biology, and other areas are often based on large data sets of thousands to millions of training instances. However, several standard learning algorithms, such as kernel-based algorithms, e.g., Support Vector Machines, Kernel Ridge Regression, Kernel PCA, do not easily scale to such orders of magnitude. This thesis focuses on sampling-based matrix approximation techniques that help scale kernel-based algorithms to large-scale datasets. We address several fundamental theoretical and empirical questions including:

  What approximation should be used? We discuss two common sampling-based methods, providing novel theoretical insights regarding their suitability for various applications and experimental results motivated by this theory. Our results show that one of these methods, the Nystrom method, is superior in the context of large-scale learning.

  Do these approximations work in practice? We show the effectiveness of approximation techniques on a variety of problems. In the largest study to-date for manifold learning, we use the Nystrom method to extract low-dimensional structure from high-dimensional data to effectively cluster face images. We also report good empirical results for kernel ridge regression and kernel logistic regression.

  How should we sample columns? A key aspect of sampling-based algorithms is the distribution according to which columns are sampled. We study both fixed and adaptive sampling schemes as well as a promising ensemble technique that can be easily parallelized and generates superior approximations, both in theory and in practice.

  How well do these approximations work in theory? We provide theoretical analyses of the Nystrom method to understand when this technique should be used. We present guarantees on approximation accuracy based on various matrix properties and analyze the effect of matrix approximation on actual kernel-based algorithms.

  This work has important consequences for the machine learning community since it extends to large-scale applications the benefits of kernel-based algorithms. The crucial aspect of this research, involving low-rank matrix approximation, is of independent interest within the field of numerical linear algebra.

• Ph.D. Thesis 2009 Factor Graphs for Relational Regression Chopra, Sumit Abstract | PDF

  Title: Factor Graphs for Relational Regression

  Candidate: Chopra, Sumit

  Advisor(s): LeCun, Yann

  Abstract:

  Inherent in many interesting regression problems is a rich underlying inter-sample "Relational Structure". In these problems, the samples may be related to each other in ways such that the unknown variables associated with any sample not only depends on its individual attributes, but also depends on the variables associated with related samples. One such problem, whose importance is further emphasized by the present economic crises, is understanding real estate prices. The price of a house clearly depends on its individual attributes, such as, the number of bedrooms. However, the price also depends on the neighborhood in which the house lies and on the time period in which it was sold. This effect of neighborhood and time on the price is not directly measurable. It is merely reflected in the prices of other houses in the vicinity that were sold around the same time period. Uncovering these spatio-temporal dependencies can certainly help better understand house prices, while at the same time improving prediction accuracy.

  Problems of this nature fall in the domain of "Statistical Relational Learning". However the drawback of most models proposed so far is that they cater only to classification problems. To this end, we propose "relational factor graph" models for doing regression in relational data. A single factor graph is used to capture, one, dependencies among individual variables of sample, and two, dependencies among variables associated with multiple samples. The proposed models are capable of capturing hidden inter-sample dependencies via latent variables, and also permits non-linear log-likelihood functions in parameter space, thereby allowing considerably more complex architectures. Efficient inference and learning algorithms for relational factor graphs are proposed. The models are applied to predicting the prices of real estate properties and for constructing house price indices. The relational aspect of the model accounts for the hidden spatio-temporal influences on the price of every house. Experiments show that one can achieve considerably superior performance by identifying and using the underlying spatio-temporal structure associated with the problem. To the best of our knowledge this is the first work in the direction of relational regression and is also the first work in constructing house price indices by simultaneously accounting for the spatio-temporal effects on house prices using large-scale industry standard data set.

• Ph.D. Thesis 2009 Numerical Estimation of the Second Largest Eigenvalue of a Reversible Markov Transition Matrix Gade, Kranthi Abstract | PDF

  Title: Numerical Estimation of the Second Largest Eigenvalue of a Reversible Markov Transition Matrix

  Candidate: Gade, Kranthi

  Advisor(s): Goodman, Jonathan

  Abstract:

  We discuss the problem of finding the second largest eigenvalue of an operator that defines a reversible Markov chain. The second largest eigenvalue governs the rate at which the statistics of the Markov chain converge to equilibrium. Scientific applications include understanding the very slow dynamics of some models of dynamic glass. Applications in computing include estimating the rate of convergence of Markov chain Monte Carlo algorithms.

  Most practical Markov chains have state spaces so large that direct or even iterative methods from linear algebra are inapplicable. The size of the state space, which is the dimension of the eigenvalue problem, grows exponentially with the system size. This makes it impossible to store a vector (for sparse methods), let alone a matrix (for dense methods). Instead, we seek a method that uses only time correlation from samples produced from the Markov chain itself.

  In this thesis, we propose a novel Krylov subspace type method to estimate the second eigenvalue from the simulation data of the Markov chain using test functions which are known to have good overlap with the slowest mode. This method starts with the naive Rayleigh quotient estimate of the test function and refines it to obtain an improved estimate of the second eigenvalue. We apply the method to a few model problems and the estimate compares very favorably with the known answer. We also apply the estimator to some Markov chains occuring in practice, most notably in the study of glasses. We show experimentally that our estimator is more accurate and stable for these problems compared to the existing methods.

• Ph.D. Thesis 2009 2D-Centric Interfaces and Algorithms for 3D Modeling Gingold, Yotam Abstract | PDF

  Title: 2D-Centric Interfaces and Algorithms for 3D Modeling

  Candidate: Gingold, Yotam

  Advisor(s): Zorin, Denis

  Abstract:

  The creation of 3D models is a fundamental task in computer graphics. The task is required by professional artists working on movies, television, and games, and desired by casual users who wish to make their own models for use in virtual worlds or as a hobby.

  In this thesis, we consider approaches to creating and editing 3D models that minimize the user's thinking in 3D. In particular, our approaches do not require the user to manipulate 3D positions in space or mentally invert complex 3D-to-2D mappings. We present interfaces and algorithms for the creation of 3D surfaces, for texturing, and for adding small-to-medium scale geometric detail.

  First, we present a novel approach for texture placement and editing based on direct manipulation of textures on the surface. Compared to conventional tools for surface texturing, our system combines UV-coordinate specification and texture editing into one seamless process, reducing the need for careful initial design of parameterization and providing a natural interface for working with textures directly on 3D surfaces.

  Second, we present a system for free-form surface modeling that allows a user to modify a shape by changing its rendered, shaded image using stroke-based drawing tools. A new shape, whose rendered image closely approximates user input, is c omputed using an efficient and stable surface optimization procedure. We demonstrate how several types of free-form surface edits which may be difficult to cast in terms of standard deformation approaches can be easily performed using our system.

  Third, we present a single-view 2D interface for 3D modeling based on the idea of placing 2D primitives and annotations on an existing, pre-made sketch or image. Our interface frees users to create 2D sketches from arbitrary angles using their preferred tool---including pencil and paper---which they then "describe" using our tool to create a 3D model. Our primitives are manipulated with persistent, dynamic handles, and our annotations take the form of markings commonly used in geometry textbooks.

• Ph.D. Thesis 2009 Proximity problems for point sets residing in spaces with low doubling dimension Gottlieb, Lee-Ad Abstract | PDF

  Title: Proximity problems for point sets residing in spaces with low doubling dimension

  Candidate: Gottlieb, Lee-Ad

  Advisor(s): Cole, Richard

  Abstract:

  In this thesis we consider proximity problems on point sets. Proximity problems arise in all fields of computer science, with broad application to computation geometry, machine learning, computational biology, data mining and the like. In particular, we will consider the problems of approximate nearest neighbor search, and dynamic maintenance of a spanner for a point set.

  It has been conjectured that all algorithms for these two problems suffer from the "curse of dimensionality," which means that their run time grow exponentially with the dimension of the point set. To avoid this undesirable growth, we consider point sets that occupy a doubling dimension lambda. We first present a dynamic data structure that uses linear space and supports a (1+e)-approximate nearest neighbor search of the point set. We then extend this algorithm to allow the dynamic maintenance of a low degree (1+e)-spanner for the point set. The query and update time of these structures are exponential in lambda (as opposed to exponential in the dimension); when lambda is small, this provides a significant spead-up over known algorithms, and when lambda is constant then these run times are optimal up to a constant. Even when no assumptions are made on lambda, the query and update times of the neighest neighbor search structure match the best known run times for approximate nearest neighbor search (up to a constant multiple in lambda). Further, the stretch of the spanner is optimal, and its update times exceed all previously known algorithms.

• Ph.D. Thesis 2009 Creativity Support for Computational Literature Howe, Daniel Abstract | PDF

  Title: Creativity Support for Computational Literature

  Candidate: Howe, Daniel

  Advisor(s): Perlin, Ken

  Abstract:

  The creativity support community has a long history of providing valuable tools to artists and designers. Similarly, creative digital media practice has proven a valuable pedagogical strategy for teaching core computational ideas. Neither strain of research has focused on the domain of literary art however, instead targeting visual, and aural media almost exclusively. 

  To address this situation, this thesis presents a software toolkit created specifically to support creativity in computational literature. Two primary hypotheses direct the bulk of the research presented: first, that it is possible to implement effective creativity support tools for literary art given current resource constraints; and second, that such tools, in addition to facilitating new forms of literary creativity, will provide unique opportunities for computer science education.

  Designed both for practicing artists and for pedagogy, the research presented directly addresses impediments to participation in the field for a diverse range of users and provides an end-to-end solution for courses attempting to engage the creative faculties of computer science students, and to introduce a wider demographic--from writers, to digital artists, to media and literary theorists --to procedural literacy and computational thinking.

  The tools and strategies presented have been implemented, deployed, and iteratively refined in a real-world contexts over the past three years. In addition to their use in large-scale projects by contemporary artists, they have provided effective support for multiple iterations of 'Programming for Digital Art & Literature', a successful inter-disciplinary computer science course taught by the author.

  Taken together, this thesis provides a novel set of tools for a new domain, and demonstrates their real-world efficacy in providing both creativity and pedagogical support for a diverse and emerging population of users.

• Ph.D. Thesis 2009 Efficient Systems Biology Algorithms for Biological Networks over Multiple Time-Scales: From Evolutionary to Regulatory Time Mitrofanova, Antonina Abstract | PDF

  Title: Efficient Systems Biology Algorithms for Biological Networks over Multiple Time-Scales: From Evolutionary to Regulatory Time

  Candidate: Mitrofanova, Antonina

  Advisor(s): Mishra, Bud

  Abstract:

  Recently, Computational Biology has emerged as one of the most exciting areas of computer science research, not only because of its immediate impact on many biomedical applications, (e.g., personalized medicine, drug and vaccine discovery, tools for diagnostics and therapeutic interventions, etc.), but also because it raises many new and interesting combinatorial and algorithmic questions, in the process. In this thesis, we focus on robust and efficient algorithms to analyze biological networks, primarily targeting protein networks, possibly the most fascinating networks in computational biology in terms of their structure, evolution and complexity, as well as because of their role in various genetic and metabolic diseases.

  Classically, protein networks have been studied statically, i.e., without taking into account time-dependent metamorphic changes in network topology and functionality. In this work, we introduce new analysis techniques that view protein networks as being dynamic in nature, evolving over time, and diverse in regulatory patterns at various stages of the system development. Our analysis is capable of dealing with multiple time-scales: ranging from the slowest time-scale corresponding to evolutionary time between species, speeding up to inter-species pathway evolution time, and finally, moving to the other extreme at the cellular developmental time-scale.

  We also provide a new method to overcome limitations imposed by corrupting effects of experimental noise (e.g., high false positive and false negative rates) in Yeast Two-Hybrid (Y2H) networks, which often provide primary data for protein complexes. Our new combinatorial algorithm measures connectivity between proteins in Y2H network not by edges but by edge-disjoint paths, which reflects pathway evolution better within single specie network. This algorithm has been shown to be robust against increasing false positives and false negatives, as estimated using variation of information and separation measures.

  In addition, we have devised a new way to incorporate evolutionary information in order to significantly improve classification of proteins, especially those isolated in their own networks or surrounded by poorly characterized neighbors. In our method, the networks of two (or more) species are joined by edges of high sequence similarity so that protein-homologs of different species can exchange information and acquire new and improved functional associations.

  Finally, we have integrated many of these techniques into one tool to create a novel analysis of malaria parasite P. falciparum's life-cycle at the scale of reaction-time, single cell level, and encompassing its entire inter-erythrocytic developmental cycle (IDC). Our approach allows connecting time-course gene expression profiles of consecutive IDC stages in order to assign functions to un-annotated Malaria proteins and predict potential targets for vaccine and drug development.

• Ph.D. Thesis 2009 Detecting, modeling and rendering complex configurations of curvilinear features Parilov, Evgueni Abstract | PDF

  Title: Detecting, modeling and rendering complex configurations of curvilinear features

  Candidate: Parilov, Evgueni

  Advisor(s): Zorin, Denis

  Abstract:

  Curvilinear features allow one to represent a variety of real world regular patterns like honeycomb tiling as well as very complicated random patterns like networks of furrows on the surface of the human skin. We have developed a set of methods and new data representations for solving key problems related to curvilinear features, which include robust detection of intricate networks of curvilinear features from digital images, GPU-based sharp rendering of fields with curvilinear features, and a parametric synthesis approach to generate systems of curvilinear features with desirable local configurations and global control.

  The existing edge-detection techniques may underperform in the presence of noise, usually do not link the detected edge points into chains, often fail on complex structures, heavily depend on initial guess, and assume significant manual phase. We have developed a technique based on active contours, or snakes, which avoids manual initial positioning of the snakes and can detect large networks of curves with complex junctions without user guidance.

  The standard bilinear interpolation of piecewise continuous fields results in unwanted smoothing along the curvilinear discontinuities. Spatially varying features can be best represented as a function of the distance to the discontinuity curves and its gradient. We have developed a real-time, GPU-based method for unsigned distance function field and its gradient field interpolation which preserves discontinuity feature curves, represented by quadratic Bezier curves, with minimal restriction on their topology.

  Detail features are very important visual clues which make computer-generated imagery look less artificial. Instead of using sample-based synthesis technique which lacks user control on features usually producing gaps in features or breaking feature coherency, we have explored an alternative approach of generating features using random fibre processes. We have developed a Gibbs-type random process of linear fibres based on local fibre interactions. It allows generating non-stationary curvilinear networks with some degree of regularity, and provides an intuitive set of parameters which directly defines fibre local configurations and global pattern of fibres.

  For random systems of linear fibres which approximately form two orthogonal dominant orientation fields, we have adapted a streamline placement algorithm which converts such systems into overlapping random sets of coherent smooth curves.

• Ph.D. Thesis 2009 Unsupervised Learning of Feature Hierarchies Ranzato, Marc'Aurelio Abstract | PDF

  Title: Unsupervised Learning of Feature Hierarchies

  Candidate: Ranzato, Marc'Aurelio

  Advisor(s): LeCun, Yann

  Abstract:

  The applicability of machine learning methods is often limited by the amount of available labeled data, and by the ability (or inability) of the designer to produce good internal representations and good similarity measures for the input data vectors.

  The aim of this thesis is to alleviate these two limitations by proposing algorithms to learn good internal representations, and invariant feature hierarchies from unlabeled data. These methods go beyond traditional supervised learning algorithms, and rely on unsupervised, and semi-supervised learning.

  In particular, this work focuses on ''deep learning'' methods, a set of techniques and principles to train hierarchical models. Hierarchical models produce feature hierarchies that can capture complex non-linear dependencies among the observed data variables in a concise and efficient manner. After training, these models can be employed in real-time systems because they compute the representation by a very fast forward propagation of the input through a sequence of non-linear transformations.

  When the paucity of labeled data does not allow the use of traditional supervised algorithms, each layer of the hierarchy can be trained in sequence starting at the bottom by using unsupervised or semi-supervised algorithms. Once each layer has been trained, the whole system can be fine-tuned in an end-to-end fashion. We propose several unsupervised algorithms that can be used as building block to train such feature hierarchies. We investigate algorithms that produce sparse overcomplete representations and features that are invariant to known and learned transformations. These algorithms are designed using the Energy-Based Model framework and gradient-based optimization techniques that scale well on large datasets. The principle underlying these algorithms is to learn representations that are at the same time sparse, able to reconstruct the observation, and directly predictable by some learned mapping that can be used for fast inference in test time.

  With the general principles at the foundation of these algorithms, we validate these models on a variety of tasks, from visual object recognition to text document classification and retrieval.

• Ph.D. Thesis 2009 Search Problems for Speech and Audio Sequences Weinstein, Eugene Abstract | PDF

  Title: Search Problems for Speech and Audio Sequences

  Candidate: Weinstein, Eugene

  Advisor(s): Mohri, Mehryar

  Abstract:

  The modern proliferation of very large audio, video, and biological databases has created a need for the design of effective methods for indexing and searching highly variable or uncertain data. Classical search and indexing algorithms deal with clean or perfect input sequences. However, an index created from speech transcriptions is marked with errors and uncertainties stemming from the use of imperfect statistical models in the speech recognition process. Similarly, automatic transcription of music, such as assigning a sequence of notes to represent a stream of music audio, is prone to errors. How can we generalize search and indexing algorithms to deal with such uncertain inputs?

  This thesis presents several novel algorithms, analyses, and general techniques and tools for effective indexing and search that not only tolerate but actually exploit this uncertainty. In particular, it develops an algorithmic foundation for music identification, or content-based music search; presents novel automata-theoretic results applicable generally to a variety of search and indexing tasks; and describes new algorithms for topic segmentation, or automatic splitting of speech streams into topic-coherent segments.

  We devise a new technique for music identification in which each song is represented by a distinct sequence of music sounds, called "music phonemes." In our approach, we learn the set of music phonemes, as well as a unique sequence of music phonemes characterizing each song, from training data using an unsupervised algorithm. We also propose a novel application of factor automata to create a compact mapping of music phoneme sequences to songs. Using these techniques, we construct an efficient and robust music identification system for a large database of songs.

  We further design new algorithms for compact indexing of uncertain inputs based on suffix and factor automata and give novel theoretical guarantees for their space requirements. Suffix automata and factor automata represent the set of all suffixes or substrings of a set of strings, and are used in numerous indexing and search tasks, including the music identification system just mentioned. We show that the suffix automaton or factor automaton of a set of strings U has at most 2Q-2 states, where Q is the number of nodes of a prefix-tree representing the strings in U, a significant improvement over previous work. We also describe a matching new linear-time algorithm for constructing the suffix automaton S or factor automaton F of U in time O(|S|).

  We also define a new quality measure for topic segmentation systems and design a discriminative topic segmentation algorithm for speech inputs, thus facilitating effective indexation of spoken audio collections. The new quality measure improves on previously used criteria and is correlated with human judgment of topic-coherence. Our segmentation algorithm uses a novel general topical similarity score based on word co-occurrence statistics. This new algorithm outperforms previous methods in experiments over speech and text streams. We further demonstrate that the performance of segmentation algorithms can be improved by using a lattice of competing hypotheses over the speech stream rather than just the one-best hypothesis as input.

• Ph.D. Thesis 2009 Using Application-Domain Knowledge in the Runtime Support of Multi-Experiment Computational Studies Yau, Siu-Man Abstract | PDF

  Title: Using Application-Domain Knowledge in the Runtime Support of Multi-Experiment Computational Studies

  Candidate: Yau, Siu-Man

  Advisor(s): Karamcheti, Vijay; Zorin, Denis

  Abstract:

  Multi-Experiment Studies (MESs) is a type of computational study in which the same simulation software is executed multiple times, and the result of all executions need to be aggregated to obtain useful insight. As computational simulation experiments become increasingly accepted as part of the scientific process, the use of MESs is becoming more wide-spread among scientists and engineers.

  MESs present several challenging requirements on the computing system. First, many MESs need constant user monitoring and feedback, requiring simultaneous steering of multiple executions of the simulation code. Second, MESs can comprise of many executions of long-running simulations; the sheer volume of computation can make them prohibitively long to run.

  Parallel architecture offer an attractive computing platform for MESs. Low-cost, small-scale desktops employing multi-core chips allow wide-spread dedicated local access to parallel computation power, offering more research groups an opportunity to achieve interactive MESs. Massively-parallel, high-performance computing clusters can afford a level of parallelism never seen before, and present an opportunity to address the problem of computationally intensive MESs.

  However, in order to fully leverage the benefits of parallel architectures, the traditional parallel systems' view has to be augmented. Existing parallel computing systems often treat each execution of the software as a black box, and are prevented from viewing an entire computational study as a single entity that must be optimized for.

  This dissertation investigates how a parallel system can view MESs as an end-to-end system and leverage the application-specific properties of MESs to address its requirements. In particular, the system can 1) adapt its scheduling decisions to the overall goal of an MES to reduce the needed computation, 2) simultaneously aggregate results from, and disseminate user actions to, multiple executions of the software to enable simultaneous steering, 3) store reusable information across executions of the simulation software to reduce individual run-time, and 4) adapt its resource allocation policies to the MES's properties to improve resource utilization.

  Using a test bed system called SimX and four example MESs across different disciplines, this dissertation shows that the application-aware MES-level approach can achieve multi-fold to multiple orders-of-magnitude improvements over the traditional simulation-level approach.

• Ph.D. Thesis 2009 Ensuring Correctness of Compiled Code Zaks, Ganna Abstract | PDF

  Title: Ensuring Correctness of Compiled Code

  Candidate: Zaks, Ganna

  Advisor(s): Pnueli, Amir

  Abstract:

  Traditionally, the verification effort is applied to the abstract algorithmic descriptions of the underlining software. However, even well understood protocols such as Peterson's protocol for mutual exclusion, whose algorithmic description takes only half a page, have published implementations that are erroneous. Furthermore, the semantics of the implementations can be altered by optimizing compilers, which are very large applications and, consequently, are bound to have bugs. Thus, it is highly desirable to ensure the correctness of the compiled code especially in safety critical and high-assurance software. This dissertation describes two alternative approaches that bring us closer to solving the problem.

  First, we present CoVaC - a deductive framework for proving program equivalence and its application to automatic verification of transformations performed by optimizing compilers. To leverage the existing program analysis techniques, we reduce the equivalence checking problem to analysis of one system - a cross-product of the two input programs. We show how the approach can be effectively used for checking equivalence of single-threaded programs that are structurally similar. Unlike the existing frameworks, our approach accommodates absence of compiler annotations and handles most of the classical intraprocedural optimizations such as constant folding, reassociation, common subexpression elimination, code motion, dead code elimination, branch optimizations, and others. In addition, we have developed rules for translation validation of interprocedural optimizations, which can be applied when compiler annotations are available.

  The second contribution is the pancam framework for verifying multi-threaded C programs. Pancam first compiles a multithreaded C program into optimized bytecode format. The framework relies on Spin, an existing explicit state model checker, to orchestrate the program's state space search. However, the program transitions and states are computed by the pancam bytecode interpreter. A feature of our approach is that not only pancam checks the actual implementation, but it can also check the code after compiler optimizations. Pancam addresses the state space explosion problem by allowing users to define data abstraction functions and to constrain the number of allowed context switches. We also describe a partial order reduction method that reduces context switches using dynamic knowledge computed on-the-fly, while being sound for both safety and liveness properties.

• Ph.D. Thesis 2008 Verification of Transactional Memories and Recursive Programs Cohen, Ariel Abstract | PDF

  Title: Verification of Transactional Memories and Recursive Programs

  Candidate: Cohen, Ariel

  Advisor(s): Pnueli, Amir

  Abstract:

  Transactional memory is a programming abstraction intended to simplify the synchronization of conflicting concurrent memory accesses without the difficulties associated with locks. In the first part of this thesis we provide a framework and tools that allow to formally verify that a transactional memory implementation satisfies its specification. First we show how to specify transactional memory in terms of admissible interchanges of transaction operations, and give proof rules for showing that an implementation satisfies its specification. We illustrate how to verify correctness, first using a model checker for bounded instantiations, and subsequently by using a theorem prover, thus eliminating all bounds. We provide a mechanical proof of the soundness of the verification method, as well as mechanical proofs for several implementations from the literature, including one that supports non-transactional memory accesses.

  Procedural programs with unbounded recursion present a challenge to symbolic model-checkers since they ostensibly require the checker to model an unbounded call stack. In the second part of this thesis we present a method for model-checking safety and liveness properties over procedural programs. Our method performs by first augmenting a concrete procedural program with a well founded ranking function, and then abstracting the Procedural programs with unbounded recursion present a challenge to symbolic model-checkers since they ostensibly require the checker to model an unbounded call stack. In the second part of this thesis we present a method for model-checking safety and liveness properties over procedural programs. Our method performs by first augmenting a concrete procedural program with a well founded ranking function, and then abstracting the augmented program by a finitary state abstraction. Using procedure summarization the procedural abstract program is then reduced to a finite-state system, which is model checked for the property.

• Ph.D. Thesis 2008 Learning Long-Range Vision for an Offroad Robot Hadsell, Raia Abstract | PDF

  Title: Learning Long-Range Vision for an Offroad Robot

  Candidate: Hadsell, Raia

  Advisor(s): LeCun, Yann

  Abstract:

  Teaching a robot to perceive and navigate in an unstructured natural world is a difficult task. Without learning, navigation systems are short-range and extremely limited. With learning, the robot can be taught to classify terrain at longer distances, but these classifiers can be fragile as well, leading to extremely conservative planning. A robust, high-level learning-based perception system for a mobile robot needs to continually learn and adapt as it explores new environments. To do this, a strong feature representation is necessary that can encode meaningful, discriminative patterns as well as invariance to irrelevant transformations. A simple realtime classifier can then be trained on those features to predict the traversability of the current terrain.

  One such method for learning a feature representation is discussed in detail in this work. Dimensionality reduction by learning an invariant mapping (DrLIM) is a weakly supervised method for learning a similarity measure over a domain. Given a set of training samples and their pairwise relationships, which can be arbitrarily defined, DrLIM can be used to learn a function that is invariant to complex transformations of the inputs such as shape distortion and rotation.

  The main contribution of this work is a self-supervised learning process for long-range vision that is able to accurately classify complex terrain, permitting improved strategic planning. As a mobile robot moves through offroad environments, it learns traversability from a stereo obstacle detector. The learning architecture is composed of a static feature extractor, trained offline for a general yet discriminative feature representation, and an adaptive online classifier. This architecture reduces the effect of concept drift by allowing the online classifier to quickly adapt to very few training samples without overtraining. After experiments with several different learned feature extractors, we conclude that unsupervised or weakly supervised learning methods are necessary for training general feature representations for natural scenes.

  The process was developed and tested on the LAGR mobile robot as part of a fully autonomous vision-based navigation system.

• Ph.D. Thesis 2008 Synthesizing Executable Programs from Requirements Plock, Cory Abstract | PDF

  Title: Synthesizing Executable Programs from Requirements

  Candidate: Plock, Cory

  Advisor(s): Goldberg, Benjamin

  Abstract:

  Automatic generation of correct software from requirements has long been a ``holy grail'' for system and software development. According to this vision, instead of implementing a system and then working hard to apply testing and verification methods to prove system correctness, a system is rather built correctly by construction. This problem, referred to as synthesis, is undecidable in the general case. However, by restricting the domain to decidable subsets, it is possible to bring this vision one step closer to reality.

  The focus of our study is reactive systems, or non-terminating programs that continuously receive input from an external environment and produce output responses. Reactive systems are often safety critical and include applications such as anti-lock braking systems, auto-pilots, and pacemakers. One of the challenges of reactive system design is ensuring that the software meets the requirements under the assumption of unpredictable environment input. The behavior of many of these systems can be expressed as regular languages over infinite strings, a domain in which synthesis has yielded successful results.

  We present a method for synthesizing executable reactive systems from formal requirements. The object-oriented requirements language of Live Sequence Charts (LSCs) is considered. We begin by establishing a mapping between various subsets of the language and finite-state formal models. We also consider LSCs which can express time constraints over a dense-time domain. From one of these models, we show how to formulate a winning strategy that is guaranteed to satisfy the requirements, provided one exists. The strategy is realized in the form of a controller which guides the system in choosing only non-violating behaviors. We describe an implementation of this work as an extension of an existing tool called the Play-Engine.

• Ph.D. Thesis 2008 Theory and Algorithms for Modern Machine Learning Problems and an Analysis of Markets Rastogi, Ashish Abstract | PDF

  Title: Theory and Algorithms for Modern Machine Learning Problems and an Analysis of Markets

  Candidate: Rastogi, Ashish

  Advisor(s): Cole, Richard; Mohri, Mehryar

  Abstract:

  The unprecedented growth of the Internet over the past decade and of data collection, more generally, has given rise to vast quantities of digital information, ranging from web documents and images, genomic databases to a vast array of business customer information. Consequently, it is of growing importance to develop tools and models that enable us to better understand this data and to design data-driven algorithms that leverage this information. This thesis provides several fundamental theoretical and algorithmic results for tackling such problems with applications to speech recognition, image processing, natural language processing, computational biology and web-based algorithms.

  Probabilistic automata provide an efficient and compact way to model sequence- oriented data such as speech or web documents. Measuring the similarity of such automata provides a way of comparing the objects they model, and is an essential first step in organizing this type of data. We present algorithmic and hardness results for computing various discrepancies (or dissimilarities) between probabilistic automata, including the relative entropy and the Lp distance; we also give an efficient algorithm to determine if two probabilistic automata are equivalent. In addition, we study the complexity of computing the norms of probabilistic automata.

  Organizing and querying large amounts of digitized data such as images and videos is a challenging task because little or no label information is available. This motivates transduction, a setting in which the learning algorithm can leverage unlabeled data during training to improve performance. We present novel error bounds for a family of transductive regression algorithms and validate their usefulness through experiments.

  Widespread success of search engines and information retrieval systems has led to large scale collection of rating information which is being used to provide personalized rankings. We examine an alternate formulation of the ranking problem for search engines motivated bythe requirement that in addition to accurately predicting pairwise ordering, ranking systems must also preserve the magnitude of the preferences or the difference between ratings. We present algorithms with sound theoretical properties, and verify their efficacy through experiments.

  Finally, price discovery in a market setting can be viewed as an (ongoing) learning problem. Specifically, the problem is to find and maintain a set of prices that balance supply and demand, a core topic in economics. This appears to involve complex implicit and possibly large-scale information transfers. We show that finding equilibrium prices, even approximately, in discrete markets is NP-hard and complement the hardness result with a matching polynomial time approximation algorithm.We also give a new way of measuring the quality of an approximation to equilibrium prices that is based on a natural aggregation of the dissatisfaction of individual market participants.

• Ph.D. Thesis 2008 Geometric Modeling with High Order Derivatives Tosun, Elif Abstract | PDF

  Title: Geometric Modeling with High Order Derivatives

  Candidate: Tosun, Elif

  Advisor(s): Zorin, Denis

  Abstract:

  Modeling of high quality surfaces is the core of geometric modeling. Such models are used in many computer-aided design and computer graphics applications. Irregular behavior of higher-order differential parameters of the surface (e.g. curvature variation) may lead to aesthetic or physical imperfections. In this work, we consider approaches to constructing surfaces with high degree of smoothness.

  One direction is based on a manifold-based surface definition which ensures well-defined high-order derivatives that can be explicitly computed at any point. We extend previously proposed manifold-based construction to surfaces with piecewise-smooth boundary and explore trade-offs in some elements of the construction. We show that growth of derivative magnitudes with order is a general property of constructions with locally supported basis functions and derive a lower bound for derivative growth and numerically study flexibility of resulting surfaces at arbitrary points.

  An alternative direction to using high-order surfaces is to define an approximation to high-order quantities for meshes, with high-order surface implicit. These approximations do not necessarily converge point-wise, but can nevertheless be successfully used to solve surface optimization problems. Even though fourth-order problems are commonly solved to obtain high quality surfaces, in many cases, these formulations may lead to reflection-line and curvature discontinuities. We consider two approaches to further increasing control over surface properties.

  The first approach is to consider data-dependent functionals leading to fourth-order problems but with explicit control over desired surface properties. Our fourth-order functionals are based on reflection line behavior. Reflection lines are commonly used for surface interrogation and high-quality reflection line patterns are well-correlated with high-quality surface appearance. We demonstrate how these can be discretized and optimized accurately and efficiently on general meshes.

  A more direct approach is to consider a poly-harmonic function on a mesh, such as the fourth-order biharmonic or the sixth-order triharmonic. The biharmonic and the triharmonic equations can be thought of as a linearization of curvature and curvature variation Euler-Lagrange equations respectively. We present a novel discretization for both problems based on the mixed finite element framework and a regularization technique for solving the resulting, highly ill-conditioned systems of equations. We show that this method, compared to more ad-hoc discretizations, has higher degree of mesh independence and yields surfaces of better quality.

• Ph.D. Thesis 2007 Scaling Data Servers via Cooperative Caching Annapureddy, Siddhartha Abstract | PDF

  Title: Scaling Data Servers via Cooperative Caching

  Candidate: Annapureddy, Siddhartha

  Advisor(s): Mazieres, David

  Abstract:

  In this thesis, we present design techniques -- and systems that illustrate and validate these techniques -- for building data-intensive applications over the Internet. We enable the use of a traditional bandwidth-limited server in these applications. A large number of cooperating users contribute resources such as disk space and network bandwidth, and form the backbone of such applications. The applications we consider fall in one of two categories. The first type provide user-perceived utility in proportion to the data download rates of the participants; bulk data distribution systems is a typical example. The second type are usable only when the participants have data download rates above a certain threshold; video streaming is a prime example.

  We built Shark, a distributed file system, to address the first type of applications. It is designed for large-scale, wide-area deployment, while also providing a drop-in replacement for local-area file systems. Shark introduces a novel locality-aware cooperative-caching mechanism, in which clients exploit each other's file caches to reduce load on an origin file server. Shark also enables sharing of data even when it originates from different servers. In addition, Shark clients are mutually distrustful in order to operate in the wide-area. Performance results show that Shark greatly reduces server load and reduces client-perceived latency for read-heavy workloads both in the wide and local areas.

  We built RedCarpet, a near-Video-on-Demand (nVoD) system, to address the second type of applications. nVoD allows a user to watch a video starting at any point after waiting for a small setup time. RedCarpet uses a mesh-based peero-peer (P2P) system to provide the nVoD service. In this context, we study the problem of scheduling the dissemination of chunks that constitute a video. We show that providing nVoD is feasible with a combination of techniques that include network coding, avoiding resource starvation for different chunks, and overay topology management algorithms. Our evaluation, using a simulator as well as a prototype, shows that systems that do not optimize in all these dimensions could deliver significantly worse nVoD performance.

• Ph.D. Thesis 2007 Shape Analysis by Abstraction, Augmentation, and Transformation Balaban, Ittai Abstract | PDF

  Title: Shape Analysis by Abstraction, Augmentation, and Transformation

  Candidate: Balaban, Ittai

  Advisor(s): Pnueli, Amir; Zuck, Lenore

  Abstract:

  The goal of shape analysis is to analyze properties of programs that perform destructive updates of linked structures (heaps). This thesis presents an approach for shape analysis based on program augmentation (instrumentation), predicate abstraction, and model checking, that allows for verification of safety and liveness properties (which, for sequential programs, usually corresponds to program invariance and termination).

  One of the difficulties in abstracting heap-manipulating programs is devising a decision procedure for a sufficiently expressive logic of graph properties. Since graph reachability (expressible by transitive closure) is not a first order property, the challenge is in showing that a decision procedure exists for a rich enough subset of first order logic with transitive closure.

  Predicate abstraction is in general too weak to verify liveness properties. Thus an additional issue dealt with is how to perform abstraction while retaining enough information. The method presented here is domain-neutral, and applies to concurrent programs as well as sequential ones.

• Ph.D. Thesis 2007 Democratizing Content Distribution Freedman, Michael Abstract | PDF

  Title: Democratizing Content Distribution

  Candidate: Freedman, Michael

  Advisor(s): Mazieres, David

  Abstract:

  In order to reach their large audiences, today's Internet publishers primarily use content distribution networks (CDNs) to deliver content. Yet the architectures of the prevalent commercial systems are tightly bound to centralized control, static deployments, and trusted infrastructure, inherently limiting their scope and scale to ensure cost recovery.

  To move beyond such shortcomings, this thesis contributes a number of techniques that realize cooperative content distribution. By federating large numbers of unreliable or untrusted hosts, we can satisfy the demand for content by leveraging all available resources. We propose novel algorithms and architectures for three central mechanisms of CDNs: content discovery (where are nearby copies of the client's desired resource?), server selection (which node should a client use?), and secure content transmission (how should a client download content efficiently and securely from its multiple potential sources?).

  These mechanisms have been implemented, deployed, and tested in production systems that have provided open content distribution services for more than three years. Every day, these systems answer tens of millions of client requests, serving terabytes of data to more than a million people.

  This thesis presents five systems related to content distribution. First, Coral provides a distributed key-value index that enables content lookups to occur efficiently and returns references to nearby cached objects whenever possible, while still preventing any load imbalances from forming. Second, CoralCDN demonstrates how to construct a self-organizing CDN for web content out of unreliable nodes, providing robust behavior in the face of failures. Third, OASIS provides a general-purpose, flexible anycast infrastructure, with which clients can locate nearby or unloaded instances of participating distributed systems. Fourth, as a more clean-slate design that can leverage untrusted participants, Shark offers a distributed file system that supports secure block-based file discovery and distribution. Finally, our authentication code protocol enables the integrity verification of large files on-the-fly when using erasure codes for efficient data dissemination.

  Taken together, this thesis provides a novel set of tools for building highly-scalable, efficient, and secure content distribution systems. By enabling the automated replication of data based on its popularity, we can make desired content available and accessible to everybody. And in effect, democratize content distribution.

• Ph.D. Thesis 2007 Joint Inference for Information Extraction and Translation Ji, Heng Abstract | PDF

  Title: Joint Inference for Information Extraction and Translation

  Candidate: Ji, Heng

  Advisor(s): Grishman, Ralph

  Abstract:

  The traditional natural language processing pipeline incorporates multiple stages of linguistic analysis. Although errors are typically compounded through the pipeline, it is possible to reduce the errors in one stage by harnessing the results of the other stages.

  This thesis presents a new framework based on component interactions to approach this goal. The new framework applies all stages in a suitable order, with each stage generating multiple hypotheses and propagating them through the whole pipeline. Then the feedback from subsequent stages is used to enhance the target stage by re-ranking these hypotheses, and then produce the best analysis.

  The effectiveness of this framework has been demonstrated by substantially improving the performance of Chinese and English entity extraction and Chinese-to-English entity translation. The inference knowledge includes mono-lingual interactions among information extraction stages such as name tagging, coreference resolution, relation extraction and event extraction, as well as cross-lingual interaction between information extraction and machine translation.

  Such symbiosis of analysis components allows us to incorporate information from a much wider context, spanning the entire document and even going across documents, and utilize deeper semantic analysis; it will therefore be essential for the creation of a high- performance NLP pipeline.

• Ph.D. Thesis 2007 Authentication Mechanisms for Open Distributed Systems Nicolosi, Antonio Abstract | PDF

  Title: Authentication Mechanisms for Open Distributed Systems

  Candidate: Nicolosi, Antonio

  Advisor(s): Mazieres, David; Shoup, Victor

  Abstract:

  While authentication within organizations is a well-understood problem, traditional solutions are often inadequate at the scale of the Internet, where the lack of a central authority, the open nature of the systems, and issues such as privacy and anonymity create new challenges. For example, users typically establish dozens of web accounts with independently administered services under a single password, which increases the likelihood of exposure of their credentials; users wish to receive email from anyone who is not a spammer, but the openness of the email infrastructure makes it hard to authenticate legitimate senders; users may have a rightful expectation of privacy when viewing widely-accessed protected resources such as premium website content, yet they are commonly required to present identifying login credentials, which permits tracking of their access patterns.

  This dissertation describes enhanced authentication mechanisms to tackle the challenges of each of the above settings. Specifically, the dissertation develops: 1) a remote authentication architecture that lets users recover easily in case of password compromise; 2) a social network-based email system in which users can authenticate themselves as trusted senders without disclosing all their social contacts; and 3) a group access-control scheme where requests can be monitored while affording a degree of anonymity to the group member performing the request.

  The proposed constructions combine system designs and novel cryptographic techniques to address their respective security and privacy requirements both effectively and efficiently.

• Ph.D. Thesis 2007 New Design Criteria for Hash Functions and Block Ciphers Puniya, Prashant Abstract | PDF

  Title: New Design Criteria for Hash Functions and Block Ciphers

  Candidate: Puniya, Prashant

  Advisor(s): Dodis, Yevgeniy

  Abstract:

  Cryptographic primitives, such as hash functions and block ciphers, are integral components in several practical cryptographic schemes. In order to prove security of these schemes, a variety of security assumptions are made on the underlying hash function or block cipher, such as collision-resistance, pseudorandomness etc. In fact, such assumptions are often made without much regard for the actual constructions of these primitives. In this thesis, we address this problem and suggest new, and possibly better, design criteria for hash functions and block ciphers.

  We start by analyzing the design criteria underlying hash functions. The usual design principle here involves a two-step procedure: First, come up with a heuristically-designed and ``hopefully strong'' fixed-length input construction (i.e. the compression function), then use a standard domain extension technique, usually the cascade construction, to get a construction that works for variable-length inputs. We investigate this design principle from two perspectives:

  • To instantiate the Random Oracle. We suggest modifications to existing constructions that make the resulting construction secure as a random oracle, with appropriate assumptions on the underlying compression function.
  • In general, we look for ``black-box'' fixes to existing hash functions to get secure constructions for each of the common security notions required of hash functions. We also give suggestions for appropriate modes for using existing hash functions along these lines.

  We next move on to discuss the Feistel network, which is used in the design of several popular block ciphers such as DES, Triple-DES, Blowfish etc. Currently, the celebrated result of Luby-Rackoff (and further extensions) is regarded as the theoretical basis for using this construction in block cipher design, where it was shown that a four-round Feistel network is a (strong) pseudorandom permutation (PRP) if the round functions are independent pseudorandom functions (PRFs). We study the Feistel network from two different perspectives:

  • Is there a weaker security notion for round functions, than pseudorandomness, that suffices to prove security of the Feistel network?
  • Can the Feistel network satisfy a much stronger security notion, i.e. security as an ideal cipher, under appropriate assumptions on the round functions?

  We give a positive answer to the first question and a partial positive answer to the second question. In the process, we undertake a combinatorial study of the Feistel network, that might be useful in other scenarios as well. We provide several practical applications of our results for the Feistel network.

• Ph.D. Thesis 2007 Being Lazy and Preemptive at Learning toward Information Extraction Shinyama, Yusuke Abstract | PDF

  Title: Being Lazy and Preemptive at Learning toward Information Extraction

  Candidate: Shinyama, Yusuke

  Advisor(s): Sekine, Satoshi

  Abstract:

  This thesis proposes a novel approach for exploring Information Extraction scenarios. Information Extraction, or IE, is a task aiming at finding events and relations in natural language texts that meet a user's demand. However, it is often difficult to formulate, or even define such events that satisfy both a user's need and technical feasibility. Furthermore, most existing IE systems need to be tuned for a new scenario with proper training data in advance. So a system designer usually needs to understand what a user wants to know in order to maximize the system performance, while the user has to understand how the system will perform in order to maximize his/her satisfaction.

  In this thesis, we focus on maximizing the variety of scenarios that the system can handle instead of trying to improve the accuracy of a particular scenario. In traditional IE systems, a relation is defined a priori by a user and is identified by a set of patterns that are manually crafted or acquired in advance. We propose a technique called Unrestricted Relation Discovery, which defers determining what is a relation and what is not until the very end of the processing so that a relation can be defined a posteriori. This laziness gives huge flexibility to the types of relations the system can handle. Furthermore, we use the notion of recurrent relations to measure how useful each relation is. This way, we can discover new IE scenarios without fully specifying definitions or patterns, which leads to Preemptive Information Extraction, where a system can provide a user a portfolio of extractable relations and let the user choose them.

  We used one year news articles obtained from the Web as a development set. We discovered dozens of scenarios that are similar to the existing scenarios tried by many IE systems, as well as new scenarios that are relatively novel. We have evaluated the existing scenarios with Automatic Content Extraction (ACE) event corpus and obtained reasonable performance. We believe this system will shed new light on IE research by giving various experimental IE scenarios.

• Ph.D. Thesis 2007 Constituent Parsing by Classification Turian, Joseph Abstract | PDF

  Title: Constituent Parsing by Classification

  Candidate: Turian, Joseph

  Advisor(s): Melamed, I. Dan

  Abstract:

  We present an approach to constituent parsing, which is driven by classifiers induced to minimize a single regularized objective. It is the first discriminatively-trained constituent parser to surpass the Collins (2003) parser without using a generative model. Our primary contribution is simplifying the human effort required for feature engineering. Our model can incorporate arbitrary features of the input and parse state. Feature selection and feature construction occur automatically, as part of learning. We define a set of fine-grained atomic features, and let the learner induce informative compound features. Our learning approach includes several novel approximations and optimizations which improve the efficiency of discriminative training. We introduce greedy completion, a new agenda-driven search strategy designed to find low-cost solutions given a limit on search effort. The inference evaluation function was learned accurately enough to guide the deterministic parsers to the optimal parse reasonably quickly without pruning, and thus without search errors. Experiments demonstrate the flexibility of our approach, which has also been applied to machine translation (Wellington et. al, AMTA 2006; Turian et al., NIPS 2007).

• Ph.D. Thesis 2007 Enhanced Security Models for Network Protocols Walfish, Shabsi Abstract | PDF

  Title: Enhanced Security Models for Network Protocols

  Candidate: Walfish, Shabsi

  Advisor(s): Dodis, Yevgeniy

  Abstract:

  Modeling security for protocols running in the complex network environment of the Internet can be a daunting task. Ideally, a security model for the Internet should provide the following guarantee: a protocol that "securely" implements a particular task specification will retain all the same security properties as the specification itself, even when an arbitrary set of protocols runs concurrently on the same network. This guarantee must hold even when other protocols are maliciously designed to interact badly with the analyzed protocol, and even when the analyzed protocol is composed with other protocols. The popular Universal Composability (UC) security framework aims to provide this guarantee.

  Unfortunately, such strong security guarantees come with a price: they are impossible to achieve without the use of some trusted setup. Typically, this trusted setup is global in nature, and takes the form of a Public Key Infrastructure (PKI) and/or a Common Reference String (CRS). However, the current approach to modeling security in the presence of such setups falls short of providing expected security guarantees. A quintessential example of this phenomenon is the deniability concern: there exist natural protocols that meet the strongest known security notions (including UC) while failing to provide the same deniability guarantees that their task specifications imply they should provide.

  We introduce the Generalized Universal Composability (GUC) framework to extend the UC security notion and enable the re-establishment of its original intuitive security guarantees even for protocols that use global trusted setups. In particular, GUC enables us to guarantee that secure protocols will provide the same level of deniability as the task specification they implement. To demonstrate the usefulness of the GUC framework, we first apply it to the analysis and construction of deniable authentication protocols. Building upon such deniable authentication protocols, we then prove a general feasibility result showing how to construct protocols satisfying our security notion for a large class of two-party and multi-party tasks (assuming the availability of some reasonable trusted setup). Finally, we highlight the practical applicability of GUC by constructing efficient protocols that securely instantiate two common cryptographic tasks: commitments and zero-knowledge proofs.

• Ph.D. Thesis 2007 Tree-Structured Models of Multitext: Theory, Design and Experiments Wellington, Benjamin Abstract | PDF

  Title: Tree-Structured Models of Multitext: Theory, Design and Experiments

  Candidate: Wellington, Benjamin

  Advisor(s): Melamed, I. Dan

  Abstract:

  Statistical machine translation (SMT) systems use empirical models to simulate the act of human translation between language pairs. This dissertation surveys the ability of currently popular syntax-aware SMT systems to model real-world multitext, and shows different types of linguistic phenomena occurring in natural language translation that these popular systems cannot capture. It then proposes a new grammar formalism, Generalized Multitext Grammar (GMTG), and a generalization of Chomsky Normal Form, that allows us to build an efficient SMT system using previously developed parsing techniques. The dissertation addresses many software engineering issues that arise when doing syntax-based SMT using large corpora and lays out a object-oriented design for a translation toolkit. Using the toolkit, we show that a tree-transduction based SMT system, which uses modern machine learning algorithms, outperforms a generative baseline.

• Ph.D. Thesis 2007 Formal Verification Using Static and Dynamic Analyses Zaks, Aleksandr Abstract | PDF

  Title: Formal Verification Using Static and Dynamic Analyses

  Candidate: Zaks, Aleksandr

  Advisor(s): Pnueli, Amir

  Abstract:

  One of the main challenges of formal verification is the ability to handle systems of realistic size, which is especially exacerbated in the context of software verification. In this dissertation, we suggest two related approaches that, while both rely on formal method techniques, they can still be applied to larger practical systems. The scalability is mainly achieved by restricting the types of properties we are considering and guarantees that are given.

  Our first approach is a novel run-time monitoring framework. Unlike previous work on this topic, we expect the properties to be specified using Property Specification Language (PSL). PSL is a newly adopted IEEE P1850 standard and is an extension of Linear Temporal Logic (LTL). The new features include regular expressions and finite trace semantics, which make the new logic very attractive for run-time monitoring of both software and hardware designs. To facilitate the new logic we have extended the existing algorithm for LTL tester construction to cover the PSL specific operators. Another novelty of our approach is the ability to use partial information about the program that is being monitored while the existing tools only use the information about the observed trace and the property under consideration. This allows going beyond the focus of traditional run-time monitoring tools -- error detection in the execution trace, towards the focus of static analysis -- bug detection in programs.

  In our second approach, we employ static analysis to compute SAT-based function summaries to detect invalid pointer accesses. To compute function summaries, we propose new techniques for improving the precision and performance in order to reduce the false error rates. In particular, we use BDDs to represent a symbolic simulation of functions, where BDDs allow an efficient representation of path-sensitive information and high level simplification. In addition, we use light-weight range analysis technique for determining lower and upper bounds for program variables, which can further offload the work form the SAT solver. Note that while in our current implementation the analysis happens at compile time, we can also use the function summaries as a basis for run-time monitoring.

• Ph.D. Thesis 2006 Guaranteed Precision for Transcendental and Algebraic Computation Made Easy Du, Zilin Abstract | PDF

  Title: Guaranteed Precision for Transcendental and Algebraic Computation Made Easy

  Candidate: Du, Zilin

  Advisor(s): Yap, Chee

  Abstract:

  Numerical non-robustness is a well-known phenomenon when implementing geometric algorithms. A general approach to achieve geometric robustness is Exact Geometric Computation (EGC). This dissertation explores the redesign and extension of Core Library, a C++ library which embraces the EGC approach. The contributions of this thesis are organized into three parts.

  In the first part, we discuss the redesign of Core Library, especially the expression "Expr" and bigfloat "BigFloat" classes. Our new design emphasizes extensibility in a clean and modular way. The three facilities in "Expr", filter, root bound and bigfloat, are separated into independent modules. This allows new filters, root bounds and some bigfloat substitute to be plugged in. The key approximate evaluation and precision propagation algorithms have been greatly improved. A new bigfloat system based on MPFR and interval arithmetic has been incorporated. Our benchmark shows that the redesigned Core Library typically has 5-10 times speedup. We also provide tools to facilitate extensions of "Expr" to incorporate new type of nodes, especially transcendental nodes.

  Although the Core Library was originally designed for algebraic applications, transcendental functions are needed in many applications. In the second part, we present a complete algorithm for absolute approximation of the general hypergeometric functions. It's complexity is also given. The extension of this algorithm to ``blackbox number'' is provided. A general hypergeometric function package based on our algorithm is implemented and integrated into the Core Library based on our new design.

  Brent has shown that many elementary functions, such as $\exp, \log, \sin$, etc., can be efficiently computed using the Arithmetic-Geometric Mean (AGM) based algorithm. However, he only gave an asymptotic error analysis. The constants in the Big $O(\cdot)$ notation required for implementation are unknown. We provide a non-asymptotic error analysis of the AGM algorithm and the related algorithms for logarithm and exponential functions. These algorithms have been implemented and incorporated into the Core Library.

• Ph.D. Thesis 2006 On Cryptographic Techniques for Digital Rights Management Fazio, Nelly Abstract | PDF

  Title: On Cryptographic Techniques for Digital Rights Management

  Candidate: Fazio, Nelly

  Advisor(s): Dodis, Yevgeniy

  Abstract:

  With more and more content being produced, distributed, and ultimately rendered and consumed in digital form, devising effective Content Protection mechanisms and building satisfactory Digital Rights Management (DRM) systems have become top priorities for the Publishing and Entertaining Industries.

  To help tackle this challenge, several cryptographic primitives and constructions have been proposed, including mechanisms to securely distribute data over a unidirectional insecure channel (Broadcast Encryption), schemes in which leakage of cryptographic keys can be traced back to the leaker (Traitor Tracing), and techniques to combine revocation and tracing capabilities (Trace-and-Revoke schemes).

  In this thesis, we present several original constructions of the above primitives, which improve upon existing DRM-enabling cryptographic primitives along the following two directions:

  1. Widening their scope of applicability e.g., by considering models taking into accounts usability issues typical of the DRM setting; and
  2. Strengthening their security guarantees to higher levels that are standards, for example, in the case of stand-alone encryption.

  Our results along the first line of work include the following:

  • An efficient public-key broadcast encryption scheme, which allows mutually mistrusting content providers to leverage a common delivery infrastructure, and can cope with low-end, stateless receivers;
  • A traitor tracing scheme with optimal transmission rate, in which encryption does not cause a blow-up in the size of the content, thus allowing for optimal utilization of the broadcast channel;
  • A public-key tracing and revoking scheme that can deal with both server-side and client-side scalability issues, while preserving traceability.

  As for the second direction, our contribution can be divided as follows:

  • A forward-secure public-key broadcast encryption scheme, in which the unauthorized access resulting from cracking a user-key is constrained to a minimal time frame which is delimited, in the future, by the revocation mechanism, and in the past, by forward secrecy;
  • A precise formalization of the notion of adaptive chosen-ciphertext security for public-key broadcast encryption schemes, along with a modular and efficient construction.

  Overall, the cryptographic tools developed in this thesis provide more flexibility and more security than existing solutions, and thus offer a better match for the challenges of the DRM setting.

• Ph.D. Thesis 2006 Finding Your Match: Techniques for Improving Sequence Alignment in DNA and RNA Gill, Ofer Hirsch Abstract | PDF

  Title: Finding Your Match: Techniques for Improving Sequence Alignment in DNA and RNA

  Candidate: Gill, Ofer Hirsch

  Advisor(s): Mishra, Bud

  Abstract:

  In Bioinformatics, finding correlations between species allows us the better understand the important biological functions of those species and trace its evolution. This thesis considers sequence alignment, a method for obtaining these correlations. We improve upon sequence alignment tools designed for DNA with Plains, an algorithm than uses piecewise-linear gap functions and parameter-optimization to obtain correlations in remotely-related species pairs such as human and fugu using reasonable amounts of memory and space on an ordinary computer. We then discuss Planar, which is similar to Plains, but is designed for aligning RNA, and accounts for secondary structure. We also explore SEPA, a tool that uses p-value estimation based on exhaustive empirical data to better emphasize key results from an alignment with a measure of reliability. Using SEPA to measure the quality of an alignment, we proceed to compare Plains and Planar against similar alignment tools, emphaisizing the interesting correlations caught in the process.

• Ph.D. Thesis 2006 DataSlicer: A Hosting Platform for Data-Centric Network Services He, Congchun Abstract | PDF

  Title: DataSlicer: A Hosting Platform for Data-Centric Network Services

  Candidate: He, Congchun

  Advisor(s): Karamcheti, Vijay

  Abstract:

  As the Web evolves, the number of network services deployed on the Internet has been growing at a dramatic pace. Such services usually involve a massive volume of data stored in physical or virtual back-end databases, and access the data to dynamically generate responses for client requests. These characteristics restrict use of traditional mechanisms for improving service performance and scalability: large volumes prevent replication of the service data at multiple sites required by content distribution schemes, while dynamic responses do not support the reuse required by web caching schemes.

  However, many deployed data-centric network services share other properties that can help alleviate this situation: (1) service usage patterns exhibit locality of various forms, and (2) services are accessed using standard protocols and publicly known message structures. When properly exploited, these characteristics enable the design of alternative caching infrastructures, which leverage distributed network intermediaries to inspect traffic flowing between clients and services, infer locality information dynamically, and potentially improve service performance by taking actions such as partial service replication, request redirection, or admission control.

  This dissertation investigates the nature of locality in service usage patterns for two well-known web services, and reports on the design, implementation, and evaluation of such a network intermediary architecture, named DataSlicer. DataSlicer incorporates four main techniques: (1) Service-neutral request inspection and locality detection on distributed network intermediaries; (2) Construction of oriented overlays for clustering client requests; (3)Integrated load-balancing and service replication mechanisms that improve service performance and scalability by either redistributing the underlying traffic in the network or creating partial service replicas on demand at appropriate network locations; and (4) Robustness mechanisms to maintain system stability in a wide-area network environment.

  DataSlicer has been successfully deployed on the PlanetLab network. Extensive experiments using synthetic workloads show that our approach can: (1) create appropriate oriented overlays to cluster client requests according to multiple application metrics; (2) detect locality information across multiple dimensions and granularity levels; (3) leverage the detected locality information to perform appropriate load-balancing and service replication actions with minimal cost; and (4) ensure robust behavior in the face of dynamically changing network conditions.

• Ph.D. Thesis 2006 Multimarker Genetic Analysis Methods for High Throughput Array Data Ionita, Iuliana Abstract | PDF

  Title: Multimarker Genetic Analysis Methods for High Throughput Array Data

  Candidate: Ionita, Iuliana

  Advisor(s): Mishra, Bud

  Abstract:

  In this thesis, we focus on multi-marker/-locus statistical methods for analyzing high-throughput array data used for the detection of genes implicated in complex disorders. There are two main parts: the first part concerns the localization of cancer genes from copy number variation data, with an application to lung cancer; the second part concerns the localization of disease genes using an affected-sib-pair design, with an application to inflammatory bowel disease. A third part addresses an important issue involved in the design of these disease-gene-detection studies. More details follow: 

  
  1. Detection of Oncogenes and Tumor Suppressor Genes using Multipoint Statistics from Copy Number Variation Data
  
  ArrayCGH is a microarray-based comparative genomic hybridization technique that has been used to compare a tumor genome against a normal genome, thus providing rapid genomic assays of tumor genomes in terms of copy number variations of those chromosomal segments, which have been gained or lost. When properly interpreted, these assays are likely to shed important light on genes and mechanisms involved in initiation and progression of cancer. Specifically, chromosomal segments, amplified or deleted in a group of cancer patients, point to locations of cancer genes. We describe a statistical method to estimate the location of such genes by analyzing segmental amplifications and deletions in the genomes from cancer patients and the spatial relation of these segments to any specific genomic interval. The algorithm assigns to a genomic segment a score that parsimoniously captures the underlying biology. It computes a p-value for every putative disease gene by using results from the theory of scan statistics. We have validated our method using simulated datasets, as well as a real dataset on lung cancer.
  
  2. Multi-locus Linkage Analysis of Affected-Sib-Pairs
  
  A The affected-sib-pair (ASP) design is a simple and popular design in the linkage analysis of complex traits. The traditional ASP methods evaluate the linkage information at a locus by considering only the marginal linkage information present at that locus. However complex traits are influenced by multiple genes that together interact to increase the risk to disease. We describe a multi-locus linkage method that uses both the marginal information and information derived from the possible interactions among several disease loci, thereby increasing the significance of loci with modest marginal effects. Our method is based on a statistic that quantifies the linkage information contained in a set of markers. By a marker selection-reduction process, we screen a set of polymorphisms and select a few that seem linked to disease. We test our approach on simulated data and a genome-scan data for inflammatory bowel disease. We show that our method is expected to be more powerful than single-locus methods in detecting disease loci responsible for complex traits.
  
  3. A Practical Haplotype Inference Algorithm
  
  We consider the problem of efficient inference algorithms to determine the haplotypes and their distribution from a dataset of unrelated genotypes.
  
  With the currently available catalogue of single-nucleotide polymorphisms (SNPs) and given their abundance throughout the genome (one in about $500$ bps) and low mutation rates, scientists hope to significantly improve their ability to discover genetic variants associated with a particular complex trait. We present a solution to a key intermediate step by devising a practical algorithm that has the ability to infer the haplotype variants for a particular individual from its own genotype SNP data in relation to population data. The algorithm we present is simple to describe and implement; it makes no assumption such as perfect phylogeny or the availability of parental genomes (as in trio-studies); it exploits locality in linkages and low diversity in haplotype blocks to achieve a linear time complexity in the number of markers; it combines many of the advantageous properties and concepts of other existing statistical algorithms for this problem; and finally, it outperforms competing algorithms in computational complexity and accuracy, as demonstrated by the studies performed on real data and synthetic data.

   

• Ph.D. Thesis 2006 Expressive Motion Lees, Alyssa Abstract | PDF

  Title: Expressive Motion

  Candidate: Lees, Alyssa

  Advisor(s): Bregler, Christopher; Geiger, Davi

  Abstract:

  Since the advent of motion capture animation, attempts have been made to extract the seemingly nebulously defined attributes of 'content' and 'style' from the motion data. Enabling quick access to highly precise data, the benefits of motion capture for animation purposes are abundant. Yet manipulating the expressive attributes of the motion data in a comprehensive manner has proved elusive. This dissertation poses practical solutions that are based on insights from the dance community and learning attributes from the motion data itself. The culminating project is a system which learns the deformations of the human body and reapplies them in exaggerated form for enhanced expressivity.

  While simultaneously developing efficient and usable tools for animators, the result is a three pronged technique to enhance the expressive qualities of motion capture animation. The key aspect is the creation of a deformable skeleton representation of the human body using a unique machine learning approach. The deformable skeleton is modeled by replicating the actual movements of the human spine. The second step relies on exploiting the subtle aspects of motion, such as hand movement to create an emotional effect visually. Both of these approaches involve exaggerating the movements in the same vein as traditional 2-D animation technique of 'squash and stretch'. Finally, a novel technique for the application of style on a baseline motion capture sequence is developed.

  All of these approaches are rooted in machine learning techniques. Linear discriminate analysis was initially applied to a single phrase of motion demonstrating various style characteristics in LABAN notation. A variety of methods including nonlinear PCA, and LLE were used to learn the underlying manifold of spine movements. Nonlinear dynamic models were learned in attempts to describe motion segments versus single phrases. In addition, the dissertation focuses on the variety of obstacles in learning with motion data. This includes the correct parameterization of angles, applying statistical analysis to quaternions, and appropriate distance measures between postures.

• Ph.D. Thesis 2006 Building Trustworthy Storage Services out of Untrusted Infrastructure Li, Jinyuan Abstract | PDF

  Title: Building Trustworthy Storage Services out of Untrusted Infrastructure

  Candidate: Li, Jinyuan

  Advisor(s): Mazieres, David

  Abstract:

  As the Internet has become increasingly ubiquitous, it has seen tremendous growth in the popularity of online services. These services range from online CVS repositories like sourceforge , shopping sites, to online financial and administrative systems, etc. It is critical for these services to provide correct and reliable execution for clients. However, given their attractiveness as targets and ubiquitous accessibility, online servers also have a significant chance of being compromised, leading to Byzantine failures.

  Designing and implementing a service to run on a machine that may be compromised is not an easy task, since infrastructure under malicious control may behave arbitrarily. Even worse, as any monitoring facility may also be subverted at the same time, there is no easy way for system behavior to be audited, or for malicious attacks to be detected.

  We propose our solution to the problem by reducing the trust needed on the server side in the first place. In the other words, our system is designed specifically for running on untrusted hosts. In this thesis, we realize this principle by two different approaches. First, we design and implement a new network file system -- SUNDR. In SUNDR, malicious servers cannot forge users' operations or tamper with their data without being detected. In the worst case, attackers can only conceal users' operations from each other. Still, SUNDR is able to detect this misbehavior whenever users communicate with each other directly.

  The limitation of the approach above lies in that the system cannot guarantee ideal consistency with even one single failure. In the second approach, we use replicated state machines to tolerate some fraction of malicious server failures, which is termed Byzantine Fault Tolerance (BFT) in the literature. Classical BFT systems assume less than 1/3 of the replicas are malicious, to provide ideal consistency. In this thesis, we push the boundary from 1/3 to 2/3. With fewer than 1/3 of replicas faulty, we provide the same guarantees as classical BFT systems. Additionally, we guarantee weaker consistency, instead of arbitrary behavior, when between 1/3 and 1/3 of replicas fail.

• Ph.D. Thesis 2006 Measures for Robust Stability and Controllability Mengi, Emre Abstract | PDF

  Title: Measures for Robust Stability and Controllability

  Candidate: Mengi, Emre

  Advisor(s): Overton, Michael

  Abstract:

  A linear time-invariant dynamical system is robustly stable if the system as well as all of its nearby systems in a neighborhood of interest are stable. An important property of robustly stable systems is they decay asymptotically without exhibiting significant transient behavior. The first part of this thesis work focuses on measures revealing the degree of robust stability of a dynamical system. We put special emphasis on pseudospectral measures, those based on the eigenvalues of nearby matrices for a first-order system or matrix polynomials for a higher-order system. We present algorithms for the computation of pseudospectral measures for continuous and discrete systems with quadratic rate of convergence and analyze their accuracy in the presence of rounding errors. We also provide an efficient algorithm for the numerical radius of a matrix, the modulus of the outermost point in the field of values (the set of Rayleigh quotients) of the matrix. These algorithms are inspired by algorithms of Byers, Boyd-Balakrishnan and Burke-Lewis-Overton.

  The second part is devoted to indicators of robust controllability. We call a system robustly controllable if it is controllable and remains controllable under perturbations of interest. We describe efficient methods for the computation of the distance to the closest uncontrollable system. Our first algorithm for the first-order distance to uncontrollability depends on a grid and is well-suited for low precision approximation. We then discuss algorithms for high precision approximation of the first-order distance to uncontrollability. These are based on the bisection method of Gu and the trisection variant of Burke-Lewis-Overton.

  These algorithms require the extraction of the real eigenvalues of matrices of size $O(n2)$ typically at a cost of $O(n6)$, where $n$ is the dimension of the state space. We propose a new divide-and-conquer algorithm that reduces the cost to $O(n4)$ on average in both theory and practice and $O(n5)$ in the worst case. The new iterative approach to the extraction of real eigenvalues may also be useful in other contexts. For higher-order systems we derive a singular value characterization and exploit this characterization for the computation of the higher-order distance to uncontrollability to low precision. The algorithms in this thesis assume arbitrary complex perturbations are applicable to the input system and usually require the extraction of the imaginary eigenvalues of Hamiltonian matrices (or even matrix polynomials) or the unit eigenvalues of symplectic pencils (or palindromic matrix polynomials).

• Ph.D. Thesis 2006 Algorithmic Algebraic Model Checking: Hybrid Automata & Systems Biology Mysore, Venkatesh Pranesh Abstract | PDF

  Title: Algorithmic Algebraic Model Checking: Hybrid Automata & Systems Biology

  Candidate: Mysore, Venkatesh Pranesh

  Advisor(s): Mishra, Bud

  Abstract:

  Systems Biology strives to hasten our understanding of the fundamental principles of life by adopting a systems-level approach for the analysis of cellular function and behavior. One popular framework for capturing the chemical kinetics of interacting biochemicals is Hybrid Automata. Our goal in this thesis is to aid Systems Biology research by improving the current understanding of hybrid automata, by developing techniques for symbolic rather than numerical analysis of the dynamics of biochemical networks modeled as hybrid automata, and by honing the theory to two classes of problems: kinetic mass action based simulation in genetic regulatory & signal transduction pathways, and pseudo-equilibrium simulation in metabolic networks.

  We first provide new constructions that prove that the "open" Hierarchical Piecewise Constant Derivative (HPCD) subclass is closer to the decidability and undecidability frontiers than was previously understood. After concluding that the HPCD-like classes are unsuitable for modeling chemical reactions, our quest for semi-decidable subclasses leads us to define the "semi-algebraic" subclass. This is the most expressive hybrid automaton subclass amenable to rigorous symbolic temporal reasoning. We begin with the bounded reachability problem, and then show how the dense-time temporal logic Timed Computation Tree Logic (TCTL) can be model-checked by exploiting techniques from real algebraic geometry, primarily real quantifier elimination. We also prove the undecidability of reachability in the Blum-Shub-Smale Turing Machine formalism. We then develop efficient approximation strategies by extending bisimulation partitioning, rectangular grid-based approximation, polytopal approximation and time discretization. We then develop a uniform algebraic framework for modeling biochemical and metabolic networks, also extending flux balance analysis. We present some preliminary results using a prototypical tool Tolque. It is a symbolic algebraic dense time model-checker for semi-algebraic hybrid automata, which uses Qepcad for quantifier elimination.

  The "Algorithmic Algebraic Model Checking" techniques developed in this thesis present a theoretically-grounded mathematically-sound platform for powerful symbolic temporal reasoning over biochemical networks and other semi-algebraic hybrid automata. It is our hope that by building upon this thesis, along with the development of computationally efficient parallelizable quantifier elimination algorithms and the integration of different computer algebra tools, scientific software systems will emerge that fundamentally transform the way biochemical networks (and other hybrid automata) are analyzed.

   

• Ph.D. Thesis 2006 Building an Automatic Phenotyping System of Developing Embryos Ning, Feng Abstract | PDF

  Title: Building an Automatic Phenotyping System of Developing Embryos

  Candidate: Ning, Feng

  Advisor(s): LeCun, Yann

  Abstract:

  This dissertation presents a learning-based system for the detection, identification, localization, and measurement of various sub-cellular structures in microscopic images of developing embryos. The system analyzes sequences of images obtained through DIC microscopy and detects cell nuclei, cytoplasm, and cell walls automatically. The system described in this dissertation is the key initial component of a fully automated phenotype analysis system.

  Our study primarily concerns the early stages of development of C. Elegans nematode embryos, from fertilization to the four-cell stage. The method proposed in this dissertation consists in learning the entire processing chain {\em from end to end}, from raw pixels to ultimate object categories.

  The system contains three modules: (1) a convolutional network trained to classify each pixel into five categories: cell wall, cytoplasm, nuclear membrane, nucleus, outside medium; (2) an Energy-Based Model which cleans up the output of the convolutional network by learning local consistency constraints that must be satisfied by label images; (3) A set of elastic models of the embryo at various stages of development that are matched to the label images.

  When observing normal (wild type) embryos it is possible to visualize important cellular functions such as nuclear movements and fusions, cytokinesis and the setting up of crucial cell-cell contacts. These events are highly reproducible from embryo to embryo. The events will deviate from normal behaviors when the function of a specific gene is perturbed, therefore allowing the detection of correlations between genes activities and specific early embryonic events. One important goal of the system is to automatically detect whether the development is normal (and therefore, not particularly interesting), or abnormal and worth investigating. Another important goal is to automatically extract quantitative measurements such as the migration speed of the nuclei and the precise time of cell divisions.

• Ph.D. Thesis 2006 A Polymorphic Type System and Compilation Scheme for Record Concatenation Osinski, Edward Abstract | PDF

  Title: A Polymorphic Type System and Compilation Scheme for Record Concatenation

  Candidate: Osinski, Edward

  Advisor(s): Goldberg, Benjamin

  Abstract:

  The notion of records, which are used to organize closely related groups of data so the group can be treated as a unit, and also provide access to the data within by name, is almost universally supported in programming languages. However, in virtually all cases, the operations permitted on records in statically typed languages are extremely limited. Providing greater flexibility in dealing with records, while simultaneously retaining the benefits of static type checking is a desirable goal.

  This problem has generated considerable interest, and a number of type systems dealing with records have appeared in the literature. In this work, we present the first polymorphic type system that is expressive enough to type a number of complex operations on records, including three forms of concatenation and natural join. In addition, the precise types of the records involved are inferred, to eliminate the burden of explicit type declarations. Another aspect of this problem is an efficient implementation of records and their associated operations. We also present a compilation method which accomplishes this goal.

• Ph.D. Thesis 2006 A Probabilistic Learning Approach to Attribute Value Inconsistency Resolution Pevzner, Ilya Abstract | PDF

  Title: A Probabilistic Learning Approach to Attribute Value Inconsistency Resolution

  Candidate: Pevzner, Ilya

  Advisor(s): Goldberg, Arthur

  Abstract:

  Resolving inconsistencies in data is a problem of critical practical importance. Inconsistent data arises whenever an attribute takes on multiple, inconsistent, values. This may occur when a particular entity is stored multiple times in one database, or in multiple databases that are combined.

  We investigate Attribute Value Inconsistency Resolution (AVIR), the problem of semi-automatically resolving data inconsistencies among multiple database records that describe the same person or thing.

  Our survey of the area shows that existing solutions are either limited in scope or impose a significant burden on their users. Either they do not cover all types of inconsistencies and attributes, or they require users to write or choose attribute resolution functions for each potentially conflicting attribute.

  Our ML based approach applies to all types of inconsistencies and attributes, and automatically selects appropriate resolution functions based on the conflicting data. We have invented and developed a system, that uses a set of binary features that detect data properties and relationships and resolution functions that merge data. Many such features and resolution functions have been written. The system uses supervised learning with maximum likelihood estimation to determine which function(s) to apply, based on which feature(s) fire.

  We have validated our system by comparing its error rate, decision rate and decision accuracy on a test data set to baseline values determined by a clairvoyant application of a standard approach where each potentially conflicting attribute is resolved by the best resolution function for the attribute.

• Ph.D. Thesis 2006 Animating Autonomous Pedestrians Shao, Wei Abstract | PDF

  Title: Animating Autonomous Pedestrians

  Candidate: Shao, Wei

  Advisor(s): Terzopoulos, Demetri

  Abstract:

  This thesis addresses the difficult open problem in computer graphics of autonomous human modeling and animation, specifically of emulating the rich complexity of real pedestrians in urban environments.

  We pursue an artificial life approach that integrates motor, perceptual, behavioral, and cognitive components within a model of pedestrians as highly capable individuals. Our comprehensive model features innovations in these components, as well as in their combination, yielding results of unprecedented fidelity and complexity for fully autonomous multi-human simulation in large urban environments. Our pedestrian model is entirely autonomous and requires no centralized, global control whatsoever.

  To animate a variety of natural interactions between numerous pedestrians and their environment, we represent the environment using hierarchical data structures, which efficiently support the perceptual queries of the autonomous pedestrians that drive their behavioral responses and sustain their ability to plan their actions on local and global scales.

  The animation system that we implement using the above models enables us to run long-term simulations of pedestrians in large urban environments without manual intervention. Real-time simulation can be achieved for well over a thousand autonomous pedestrians. With each pedestrian under his/her own autonomous control, the self-animated characters imbue the virtual world with liveliness, social (dis)order, and a realistically complex dynamic.

  We demonstrate the automated animation of human activity in a virtual train station, and we employ our pedestrian simulator in the context of virtual archaeology for visualizing urban social life in reconstructed archaeological sites. Our pedestrian simulator is also serving as the basis of a testbed for designing and experimenting with visual sensor networks in the field of computer vision.

• Ph.D. Thesis 2006 Complexity Analysis of Algorithms in Algebraic Computation Sharma, Vikram Abstract | PDF

  Title: Complexity Analysis of Algorithms in Algebraic Computation

  Candidate: Sharma, Vikram

  Advisor(s): Yap, Chee

  Abstract:

  Numerical computations with real algebraic numbers require algorithms for approximating and isolating real roots of polynomials. A classical choice for root approximation is Newton's method. For an analytic function on a Banach space, Smale introduced the concept of approximate zeros, i.e., points from which Newton's method for the function converges quadratically. To identify these approximate zeros he gave computationally verifiable convergence criteria called point estimates. However, in developing these results Smale assumed that Newton's method is computed exactly. For a system of $n$ homogeneous polynomials in $n+1$ variables, Malajovich developed point estimates for a different definition of approximate zero, assuming that all operations in Newton's method are computed with fixed precision. In the first half of this dissertation, we develop point estimates for these two different definitions of approximate zeros of an analytic function on a Banach space, but assume the strong bigfloat computational model of Brent, i.e., where all operations involve bigfloats with varying precision. In this model, we derive a uniform complexity bound for approximating a root of a zero-dimensional system of $n$ integer polynomials in $n$ variables. We also derive a non-asymptotic bound, in terms of the condition number of the system, on the precision required to implement the robust Newton method.

  The second part of the dissertation analyses the worst-case complexity of two algorithms for isolating real roots of a square-free polynomial with real coefficients: The Descartes method and Akritas' continued fractions algorithm. The analysis of both algorithms is based upon amortization bounds such as the Davenport-Mahler bound. For the Descartes method, we give a unified framework that encompasses both the power basis and the Bernstein basis variant of the method; we derive an $O(n(L+\log n))$ bound on the size of the recursion tree obtained by applying the method to a square-free polynomial of degree n with integer coefficients of bit-length $L$, the bound is tight for $L=\Omega(\log n)$; based upon this result we readily obtain the best known bit-complexity bound of $\wt{O}(n^4L2) $ for the Descartes method, where $\wt{O}$ means we ignore logarithmic factors. Similar worst case bounds on the bit-complexity of Akritas' algorithm were not known in the literature. We provide the first such bound, $\wt{O}(n^{12}L3)$, for a square-free integer polynomial of degree $n$ and coefficients of bit-length $L$.

• Ph.D. Thesis 2006 Pairwise Comparison between Genomic Sequences and Optical-Maps Sun, Bing Abstract | PDF

  Title: Pairwise Comparison between Genomic Sequences and Optical-Maps

  Candidate: Sun, Bing

  Advisor(s): Mishra, Bud

  Abstract:

  With the development and improvement of high throughput experimental technologies, massive amount of biological data including genomic sequences and optical-maps have been collected for various species. Comparative techniques play a central role in investigating the adaptive significance of organismal traits and revealing evolutionary relations among organisms by comparing these biological data. This dissertation presents two efficient comparative analysis tools used in comparative genomics and comparative optical-map study, respectively.

  A complete genome sequence of an organism can be viewed as its ultimate genetic map, in the sense that the heritable information are encoded within the DNA and the order of nucleotides along chromosomes is known. Comparative genomics can be applied to find functional sites by comparing genetic maps. Comparing vertebrate genomes requires efficient cross-species sequence alignment programs. The first tool introduced in this thesis is COMBAT (Clean Ordered Mer-Based Alignment Tool), a new mer-based method which can search rapidly for highly similar translated genomic sequences using the stable-marriage algorithm (SM) as an alignment filter. In experiments COMBAT is applied to comparative analysis between yeast genomes, and between the human genome and the recently published bovine genome. The homologous blocks identified by COMBAT are comparable with the alignments produced by BLASTP and BLASTZ.

  When genetic maps are not available, other genomic maps, including optical-maps, can be constructed. An optical map is an ordered enumeration of the restriction sites along with the estimated lengths of the restriction fragments between consecutive restriction sites. CAPO (Comparative Analysis and Phylogeny with Optical-Maps), introduced as a second technique in this thesis, is a tool for inferring phylogeny based on pairwise optical map comparison and bipartite graph matching. CAPO combines the stable matching algorithm with either the Unweighted Pair Group Method with Arithmetic Averaging (UPGMA) or the Neighbor-Joining (NJ) method for constructing phylogenetic trees. This new algorithm is capable of constructing phylogenetic trees in logarithmic steps and performs well in practice. Using optical maps constructed in silico and in vivo, our work shows that both UPGMA-flavored trees and the NJ-flavored trees produced by CAPO share substantial overlapping tree topology and are biologically meaningful.

• Ph.D. Thesis 2006 Exploiting Service Usage Information for Optimizing Server Resource Management Totok, Alexander Abstract | PDF

  Title: Exploiting Service Usage Information for Optimizing Server Resource Management

  Candidate: Totok, Alexander

  Advisor(s): Karamcheti, Vijay

  Abstract:

  It is difficult to provision and manage modern component-based Internet services so that they provide stable quality-of-service (QoS) guarantees to their clients, because: (1) component middleware are complex software systems that expose several independently tuned configurable application runtime policies and server resource management mechanisms; (2) session-oriented client behavior with complex data access patterns makes it hard to predict what impact tuning these policies and mechanisms has on application behavior; (3) component-based Internet services exhibit complex structural organization with requests of different types accessing different components and data sources, which could be distributed and/or replicated for failover, performance, or business purposes.

  This dissertation attempts to alleviate this situation by targeting three interconnected goals: (1) providing improved QoS guarantees to the service clients, (2) optimizing server resource utilization, and (3) providing application developers with guidelines for natural application structuring, which enable efficient use of the proposed mechanisms for improving service performance. Specifically, we explore the thesis that exposing and using detailed information about how clients use component-based Internet services enables mechanisms that achieve the range of goals listed above. To validate this thesis we show its applicability to the following four problems: (1) maximizing reward brought by Internet services, (2) optimizing utilization of server resource pools, (3) providing session data integrity guarantees, and (4) enabling service distribution in wide-area environments.

  The techniques that we propose for the identified problems are applicable at both the application structuring stage and the application operation stage, and range from automatic (i.e., performed by middleware in real time) to manual (i.e., involve the programmer, or the service provider). These techniques take into account service usage information exposed at different levels, ranging from high-level structure of user sessions to low level information about data access patterns and resource utilization by requests of different types. To show the benefits of the proposed techniques, we implement various middleware mechanisms in the JBoss application server, which utilizes the J2EE component model, and comprehensively evaluate them on several publicly-available sample J2EE applications - Java Pet Store, RUBiS, and our own implementation of the TPC-W web transactional benchmark. Our experimental results show that the proposed techniques achieve optimal utilization of server resources and improve application performance by up to two times for centralized Internet services and by up to 6 times for distributed ones.

• Ph.D. Thesis 2006 Time Series Matching: A Multi-Filter Approach Wang, Zhihua Abstract | PDF

  Title: Time Series Matching: A Multi-Filter Approach

  Candidate: Wang, Zhihua

  Advisor(s): Shasha, Dennis

  Abstract:

  Data arriving in time order (time series) arises in disciplines ranging from music to meteorology to finance to motion capture data, to name a few. In many cases, a natural way to query the data is what we call time series matching - a user enters a time series by hand, keyboard or voice and the system finds "similar" time series.

  Existing time series similarity measures, such as DTW (Dynamic Time Warping), can accommodate certain timing errors in the query and perform with high accuracy on small databases. However, they all have high computational complexity and the accuracy dramatically drops when the data set grows. More importantly, there are types of errors that cannot be captured by a single similarity measure.

  Here we present a general time series matching framework. This framework can easily optimize, combine and test different features to execute a fast similarity search based on the application's requirement. Basically we use a multi-filter chain and boosting algorithms to compose a ranking algorithm. Each filter is a classifier which removes bad candidates by comparing certain features of the time series data. Some filters use a boosting algorithm to combine a few different weak classifiers into a strong classifier. The final filter will give a ranked list of candidates in the reference data which matches the query data.

  The framework is applied to build query algorithms for a Query-by-Humming system. Experiments show that the algorithm has a more accurate similarity measure and its response time increases much slower than the pure DTW algorithm when the number of songs in the database increases from 60 to 1400.

• Ph.D. Thesis 2006 Incremental Web Search: Tracking Changes in the Web Wang, Ziyang Abstract | PDF

  Title: Incremental Web Search: Tracking Changes in the Web

  Candidate: Wang, Ziyang

  Advisor(s): Davis, Ernest

  Abstract:

  A large amount of new information is posted on the Web every day. Large-scale web search engines often update their index slowly and are unable to present such information in a timely manner. Here we present our solutions of searching new information from the web by tracking the changes of web documents.

  First, we present the algorithms and techniques useful for solving the following problems: detecting web pages that have changed, extracting changes from different versions of a web page, and evaluating the significance of web changes. We propose a two-level change detector: MetaDetector and ContentDetector. The combined detector successfully reduces network traffic by about 67%. Our algorithm for extracting web changes consists of three steps: document tree construction, document tree encoding and tree matching. It has linear time complexity and extracts effectively the changed content from different versions of a web page. In order to evaluate web changes, we propose a unified ranking framework combining three metrics: popularity ranking, content-based ranking and evolution ranking. Our methods can identify and deliver important new information in a timely manner.

  Second, we present an application using the techniques and algorithms we developed, named "Web Daily News Assistant (WebDNA): finding what's new on Your Web". It is a search tool that helps community users search new information on their community web. Currently WebDNA is deployed on the New York University web site.

  Third, we model the changes of web documents using survival analysis. Modeling web changes is useful for web crawler scheduling and web caching. Currently people model changes to web pages as a Poisson Process, and use a necessarily incomplete detection history to estimate the true frequencies of changes. However, other features that can be used to predict change frequency have not previously been studied. Our analysis shows that PageRank value is a good predictor. Statistically, the change frequency is a function proportional to $\exp[0.36\cdot (\ln(PageRank)+C)]$. We further study the problem of combining the predictor and change history into a unified framework. An improved estimator of change frequency is presented, which successfully reduces the error by 27.3% when the change history is short.

• Ph.D. Thesis 2006 Fast Algorithms for Burst Detection Zhang, Xin Abstract | PDF

  Title: Fast Algorithms for Burst Detection

  Candidate: Zhang, Xin

  Advisor(s): Shasha, Dennis

  Abstract:

  Events occur in every aspect of our lives. 

  An unexpectedly large number of events occurring within some certain measurement (e.g. within some time duration or a spatial region) is called a {\em burst}, suggesting unusual behaviors or activities. Bursts come up in many natural and social processes. It is a challenging task to monitor the occurrence of bursts whose lasting duration is unknown in a fast data stream environment.

  This work describes efficient data structures and algorithms for high performance burst detection under different settings. Our view is that bursts, as an unusual phenomenon, constitute a useful preliminary primitive in a knowledge discovery hierarchy. Our intent is to build a high performance primitive detection algorithm to support high-level data mining tasks.

  The work starts with an algorithmic framework including a family of data structures and a heuristic optimization algorithm to choose an efficient data structure given the inputs. The advantage of this framework is that it's adaptive to different inputs. Experiments on both synthetic data and real world data show the new framework significantly outperforms existing techniques over a variety of inputs.

  Furthermore, we present a greedy dynamic detection algorithm which handles the changing data. It evolves the structure to adapt to the incoming data. It achieves better performance in both synthetic and real data streams than a static algorithm in most cases.

  We have applied this framework to different real world applications in physics, stock trading and website traffic monitoring. All the case studies show our framework has superb performance.

  We extend this framework to multi-dimensional data and use it in an epidemiology simulation to detect infectious disease outbreak and spread.

• Ph.D. Thesis 2006 High Performance Algorithms for Multiple Streaming Time Series Zhao, Xiaojian Abstract | PDF

  Title: High Performance Algorithms for Multiple Streaming Time Series

  Candidate: Zhao, Xiaojian

  Advisor(s): Shasha, Dennis

  Abstract:

  Data arriving in time order (a data stream) arises in fields ranging from physics to finance to medicine to music, to name a few. Often the data comes from sensors (in physics and medicine for example) whose data rates continue to improve dramatically as sensor technology improves. Furthermore, the number of sensors is increasing, so analyzing data between sensors becomes ever more critical in order to distill knowledge from the data. Fast response is desirable in many applications (e.g. to aim a telescope at an activity of interest or to perform a stock trade). In applications such as finance, recent information, e.g. correlation, is of far more interest than older information, so analysis over sliding windows is a desired operation.

  These three factors -- huge data size, fast response, and windowed computation -- motivated this work. Our intent is to build a foundational library of primitives to perform online or near online statistical analysis, e.g. windowed correlation, incremental matching pursuit, burst detection, on thousands or even millions of time series. Beside the algorithms, we also propose the concept of ``uncooperative'' time series, whose power spectra are spread over all frequencies with any regularity.

  Previous work showed how to do windowed correlation with Fast Fourier Transforms and Wavelet Transforms, but such techniques don't work for uncooperative time series. This thesis will show how to use sketches (random projections) in a way that combines several simple techniques -- sketches, convolution, structured random vectors, grid structures, combinatorial design, and bootstrapping -- to achieve high performance, windowed correlation over a variety of data sets. Experiments confirm the asymptotic analysis.

  To conduct matching pursuit (MP) over time series windows, an incremental scheme is designed to reduce the computational effort. Our empirical study demonstrates a substantial improvement in speed.

  In previous work, Zhu and Shasha introduced an efficient algorithm to monitor bursts within windows of multiple sizes. We implemented it in a physical system by overcoming several practical challenges. Experimental results support the authors' linear running time analysis.

• Ph.D. Thesis 2006 Distribution of Route-Impacting Control Information in a Publish/Subscribe System with Delivery Guarantees Zhao, Yuanyuan Abstract | PDF

  Title: Distribution of Route-Impacting Control Information in a Publish/Subscribe System with Delivery Guarantees

  Candidate: Zhao, Yuanyuan

  Advisor(s): Kedem, Zvi

  Abstract:

  Event-driven middleware is a popular infrastructure for building large-scale asynchronous distributed systems. Content-based publish/subscribe systems are a type of event-driven middleware that provides service flexibility and specification expressiveness, creating opportunities for improving reliability and efficiency of the system.

  The use of route-impacting control information, such as subscription filters and access control rules, has the potential to enable efficient routing for applications that require selective and regional distribution of events. Such applications range from financial information systems to sensor networks to service-oriented architectures. However, it has been a great challenge to design correct and efficient protocols for distributing control information and exploiting it to achieve efficient and highly available message routing.

  In this dissertation, we study the problem of distributing and utilizing route-impacting control information. We present an abstract model of content-based routing and reliable delivery in redundant broker networks. Based on this model, we design a generic algorithm that propagates control information and performs content-based routing and delivers events reliably. The algorithm is efficient and light-weight in that it does not require heavy-weight consensus protocols between redundant brokers. We extend this generic algorithm to support consolidation and merging of control information. Existing protocols can be viewed as particular encodings and optimizations of the generic algorithm. We show an encoding using virtual time vectors that supports reliable delivery and deterministic dynamic access control in redundant broker networks. In our system, the semantics of reliable delivery is clearly defined even if subscription information and access control policy can dynamically change. That is, one or more subscribers of same principal will receive exactly the same sequence of messages (modulo subscription filter differences) regardless of where they are connected and the network latency and failure conditions in their parts of the network.

  We have implemented these protocols in a fully-functioning content-based publish/subscribe system - Gryphon. We evaluate its efficiency, scalability and high availability.

• Ph.D. Thesis 2005 Translation Validation of Optimizing Compilers Fang, Yi Abstract | PDF

  Title: Translation Validation of Optimizing Compilers

  Candidate: Fang, Yi

  Advisor(s): Pnueli, Amir; Zuck, Lenore

  Abstract:

  There is a growing awareness, both in industry and academia, of the crucial role of formally verifying the translation from high-level source-code into low-level object code that is typically performed by an optimizing comiler. Formally verifying an optimizing compiler, as one woule verify any other large program, is not feasible due to its size, ongoing evolution and modification, and possibly, proprietary considerations. Translation validation is a novel approach that offers an alternative to the verification of translator in general and compilers in particular: Rather than verifying the compiler itself, one constructs a validation tool which, after every run of the compiler, formally confirms that the target code produced in the run is a correct translation of the source program. This thesis work takes an important step towards ensuring an extremely high level of confidence in compilers targeted at EPIC architectures.

  In this thesis, we focus on the translation validation of structure preserving optimizations, i.e. transformations that do not modify programs' structure in a major way. This category of optimizations covers most of the global optimizations performed by compilers. This thesis has two main parts. One develops a proof rule that formally establishes the correctness of structure preserving transformation based on computational induction. The other part is the development of a tool that applies the proof rule to the automatic validation of global optimizaitons performed by Intel's ORC compiler for IA-64 architecture. With minimal instrumentation from the compiler, the tool constructs ''verification conditions'' -- formal theorems that, if valid, establish the correctness of a translation. The verificaiton conditions are then transferred to an automatic theorem prover that checks their validity. Together, the tool offers a fully automatic method to formally establish the correctness of each translation.

• Ph.D. Thesis 2005 Translation Validation of Loop Optimizations Hu, Ying Abstract | PDF

  Title: Translation Validation of Loop Optimizations

  Candidate: Hu, Ying

  Advisor(s): Goldberg, Benjamin; Barrett, Clark

  Abstract:

  Formal verification is important in designing reliable computer systems. For a critical software system, it is not enough to have a proof of correctness for the source code, there must also be an assurance that the compiler produces a correct translation of the source code into the target machine code. Verifying the correctness of modern optimizing compilers is a challenging task because of their size, their complexity, and their evolution over time. 

  In this thesis, we focus on the Translation Validation of loop optimizations. In order to validate the optimizations performed by the compiler, we try to prove the equivalence of the intermediate codes before and after the optimizations. There were previously a set of proof rules for building the equivalence relation between two programs. However, they cannot validate some cases with legal loop optimizations. We propose new proof rules to consider the conditions of loops and possible elimination of some loops, so that those cases can also be handled. According to these new proof rules, algorithms are designed to apply them to an automatic validation process.

  Based on the above proof rules, we implement an automatic validation tool for loop optimizations which analyzes the loops, guesses what kinds of loop optimizations occur, proves the validity of a combination of loop optimizations, and synthesizes a series of intermediate codes. We integrate this new loop tool into our translation validation tool TVOC, so that TVOC handles not only optimizations which do not significantly change the structure of the code, but also loop optimizations which do change the structure greatly. With this new part, TVOC has succeeded in validating many examples with loop optimizations. 

  Speculative optimizations are the aggressive optimizations that are only correct under certain conditions that cannot be known at compile time. In this thesis, we present the theory and algorithms for validating speculative optimizations and generating the runtime tests necessary for speculative optimizations. We also provide several examples and the results of the algorithms for speculative optimizations.

• Ph.D. Thesis 2005 Construction of Component-Based Applications by Planning Kichkaylo, Tatiana Abstract | PDF

  Title: Construction of Component-Based Applications by Planning

  Candidate: Kichkaylo, Tatiana

  Advisor(s): Karamcheti, Vijay; Ernest Davis

  Abstract:

  Many modern wide-area distributed systems are component-based. This approach provides great flexibility in adapting applications to the changing state of the environment and user requirements, but increases the complexity of configuring the applications. Because of the scale and heterogeneity of modern wide-area environments, manual configuration is hard, inefficient, suboptimal, and error-prone. Automated application configuration is desired.

  Constructing distributed applications requires choosing a set of components that will constitute the application instance and assigning network resources to component executions and data transfers. Stated this way, the application configuration problem (ACP) is similar to the planning (action selection) and scheduling (resource allocation) problems studied by the Artificial Intelligence (AI) community.

  This thesis investigates the problem of solving the ACP using AI planning techniques. However, the ACP poses several challenges not usually encountered and addressed by the traditional AI solutions. The problem specification for the ACP can be much larger than the solution, with the relevant portions only identified during the search. Additionally, the interactions between planning operators are numeric rather than logical. Finally, it is desirable to be able to trade off quality of the solution versus search time.

  We show that the ACP is undecidable in general. Therefore, instead of a single algorithm, we propose a set of techniques that can be used to compose an algorithm for a particular variety of the ACP that can exploit natural restrictions exhibited by that variety. These techniques address the challenges above by dynamically obtaining portions of the problem specification as necessary during the search, using envelope hierarchies based on numeric information for pruning and search guidance, and discretizing continuous variables to approximate numeric parameters without restricting the form of supported numeric functions.

  We illustrate these techniques by describing their use in algorithms tailored for two specific varieties of the ACP --- snapshot configurations for dynamic component-based frameworks, and scheduling of grid workflows with replica selection and explicit resource reservations. Experimental evaluation of the performance of these two algorithms shows that the techniques successfully achieve their goals, with acceptable run-time overhead.

• Ph.D. Thesis 2005 Extensible MultiModal Environment Toolkit (EMMET): A Toolkit for Prototyping and Remotely Testing Speech and Gesture Based Multimodal Interfaces Robbins, Christopher A. Abstract | PDF

  Title: Extensible MultiModal Environment Toolkit (EMMET): A Toolkit for Prototyping and Remotely Testing Speech and Gesture Based Multimodal Interfaces

  Candidate: Robbins, Christopher A.

  Advisor(s): Perlin, Ken

  Abstract:

  Ongoing improvements to the performance and accessibility of less conventional input modalities such as speech and gesture recognition now provide new dimensions for interface designers to explore. Yet there is a scarcity of commercial applications which utilize these modalities either independently or multimodally. This scarcity partially results from a lack of development tools and design guidelines to facilitate the use of speech and gesture.

  An integral aspect of the user interface design process is the ability to easily evaluate various design solutions through an iterative process of prototyping and testing. Through this process guidelines emerge that aid in the design of future interfaces. Today there is no shortage of tools supporting the development of conventional interfaces. However there do not exist resources allowing interface designers to easily prototype and quickly test, via remote distribution, interface designs utilizing speech and gesture.

  The thesis work for this dissertation explores the development of an Extensible MultiModal Environment Toolkit (EMMET) for prototyping and remotely testing speech and gesture based multimodal interfaces to three-dimensional environments. The overarching goals for this toolkit are to allow its users to: explore speech and gesture based interface design without requiring an understanding of the details involved in the low-level implementation of speech or gesture recognition, quickly distribute their multimodal interface prototypes via the Web, and receive multimodal usage statistics collected remotely after each use of their application.

  EMMET ultimately contributes to the field of multimodal user interface design by providing an environment to existing user interface developers in which speech and gesture recognition have been seamlessly integrated into their palette of user input options. Such seamless integration serves to increase the utilization within applications of speech and gesture modalities by removing any actual or perceived deterrents to the use of these modalities versus the use of conventional modalities. EMMET additionally strives to improve the quality of speech and gesture based interfaces by supporting the prototype-and-test development cycle through its Web distribution and usage statistics collection capabilities. These capabilities also allow developers to realize new design guidelines specific to the use of speech and gesture.

• Ph.D. Thesis 2005 Pattern Discovery for Hypotheses Generation in Biology Tsirigos, Aristotelis Abstract | PDF

  Title: Pattern Discovery for Hypotheses Generation in Biology

  Candidate: Tsirigos, Aristotelis

  Advisor(s): Shasha, Dennis

  Abstract:

  In recent years, the increase in the amounts of available genomic as well as gene expression data has provided researchers with the necessary information to train and test various models of gene origin, evolution, function and regulation. In this thesis, we present novel solutions to key problems in computational biology that deal with nucleotide sequences (horizontal gene transfer detection), amino-acid sequences (protein sub-cellular localization prediction), and gene expression data (transcription factor - binding site pair discovery). Different pattern discovery techniques are utilized, such as maximal sequence motif discovery and maximal itemset discovery, and combined with support vector machines in order to achieve significant improvements against previously proposed methods.

• Ph.D. Thesis 2005 Automatic Verification of Parameterized Systems Xu, Jiazhao Abstract | PDF

  Title: Automatic Verification of Parameterized Systems

  Candidate: Xu, Jiazhao

  Advisor(s): Pnueli, Amir

  Abstract:

  Verification plays an indispensable role in designing reliable computer hardware and software systems. With the fast growth in design complexity and the quick turnaround in design time, formal verification has become an increasingly important technology for establishing correctness as well as for finding difficult bugs. Since there is no ``silver-bullet'' to solve all verification problems, a spectrum of powerful techniques in formal verification have been developed to tackle different verification problems and complexity issues. Depending on the nature of the problem whose most salient components are the system implementation and the property specification, a proper methodology or a combination of different techniques is applied to solve the problem.

  In this thesis, we focus on the research and development of formal methods to uniformly verify parameterized systems. A parameterized system is a class of systems obtained by instantiating the system parameters. Parameterized verification seeks a single correctness proof of a property for the entire class. Although the general parameterized verification problem is undecidable [AK86], it is possible to solve special classes by applying a repertoire of techniques and heuristics. Many methods in parameterized verification require a great deal of human interaction. This makes the application of these methods to real world problems infeasible. Thus, the main focus of this research is to develop techniques that can be automated to deliver proofs of safety and liveness properties.

  Our research combines various formal techniques such as deductive methods, abstraction and model checking. One main result in this thesis is an automatic deductive method for parameterized verification. We apply small model properties of Bounded Data Systems (a special type of parameterized system) to help prove deductive inference rules for the safety properties of BDS systems. Another methodology we developed enables us to prove liveness properties of parameterized systems via an automatic abstraction method called counter abstraction . There are several useful by-products from our research: A set of heuristics is established for the automatic generation of program invariants which can benefit deductive verification in general; also we proposed methodologies for the automatic abstraction of fairness conditions that are crucial for proving liveness properties.

• Ph.D. Thesis 2005 Mobility, Route Caching, and TCP Performance in Mobile Ad Hoc Networks Yu, Xin Abstract | PDF

  Title: Mobility, Route Caching, and TCP Performance in Mobile Ad Hoc Networks

  Candidate: Yu, Xin

  Advisor(s): Johnson, David B.

  Abstract:

  In a mobile ad hoc network, mobile nodes communicate with each other through wireless links. Mobility causes frequent topology changes. This thesis addresses the fundamental challenges mobility presents to on-demand routing protocols and to TCP.

  On-demand routing protocols use route caches to make routing decisions. Due to mobility, cached routes easily become stale. To address the cache staleness issue, prior work used adaptive timeout mechanisms. However, heuristics cannot accurately estimate timeouts because topology changes are unpredictable. I propose to proactively disseminate the broken link information to the nodes that have cached the link. I define a new cache structure called a cache table to maintain the information necessary for cache updates, and design a distributed cache update algorithm. This algorithm is the first work that proactively updates route caches in an adaptive manner. Simulation results show that proactive cache updating is more efficient than adaptive timeout mechanisms. I conclude that proactive cache updating is key to the adaptation of on-demand routing protocols to mobility.

  TCP does not perform well in mobile ad hoc networks. Prior work provided link failure feedback to TCP so that it can avoid invoking congestion control mechanisms for packet losses caused by route failures. Simulation results show that my cache update algorithm significantly improves TCP throughput since it reduces the effect of mobility on TCP. TCP still suffers from frequent data and ACK losses. I propose to make routing protocols aware of lost TCP packets and help reduce TCP timeouts. I design two mechanisms that exploit cross-layer information awareness: early packet loss notification (EPLN) and best-effort ACK delivery (BEAD). EPLN notifies TCP senders about lost data. BEAD retransmits ACKs at intermediate nodes or at TCP receivers. Simulation results show that the two mechanisms significantly improve TCP throughput. I conclude that cross-layer information awareness is key to making TCP efficient in the presence of mobility.

  I also study the impact of route caching strategies on the scalability of on-demand routing protocols with mobility. I show that making route caches adapt quickly and efficiently to topology changes is key to the scalability of on-demand routing protocols with mobility.

• Ph.D. Thesis 2005 Information Extraction from Multiple Syntactic Sources Zhao, Shubin Abstract | PDF

  Title: Information Extraction from Multiple Syntactic Sources

  Candidate: Zhao, Shubin

  Advisor(s): Grishman, Ralph

  Abstract:

  Information Extraction is the automatic extraction of facts from text, which includes detection of named entities, entity relations and events. Conventional approaches to Information Extraction try to find syntactic patterns based on deep processing of text, such as partial or full parsing. The problem these solutions have to face is that as deeper analysis is used, the accuracy of the result decreases, and one cannot recover from the induced errors. On the other hand, lower level processing is more accurate and it can also provide useful information. However, within the framework of conventional approaches, this kind of information can not be efficiently incorporated.

  This thesis describes a novel supervised approach based on kernel methods to address these issues. In this approach customized kernels are used to match syntactic structures produced from different preprocessing phases. Using properties of a kernel, individual kernels are combined into composite kernels to integrate and extend all the information. The composite kernels can be used with various classifiers, such as Nearest Neighbor or Support Vector Machines (SVM). The main classifier we propose to use is SVM due to its ability to generalize in large dimensional feature spaces. We will show that each level of syntactic information can contribute to IE tasks, and low level information can help to recover from errors in deep processing.

  The new approach has demonstrated state-of-the-art performance on two benchmark tasks. The first task is detecting slot fillers for management succession events (MUC-6). For this task two types of kernels were designed, a surface kernel based on word n-grams and a kernel built on sentence dependency trees; the second task is the ACE RDR evaluation, which is to recognize relations between entities in text from newswire and broadcast news transcript. For this task, five kernels were built to represent information from sentence tokenization, syntactic parsing and dependency parsing. Experimental results for the two tasks will be shown and discussed.

• Ph.D. Thesis 2004 Partitionable Services Framework: Seamless Access to Distributed Applications Ivan, Anca Abstract | PDF

  Title: Partitionable Services Framework: Seamless Access to Distributed Applications

  Candidate: Ivan, Anca

  Advisor(s): Karamcheti, Vijay

  Abstract:

  A key problem in contemporary distributed systems is how to satisfy user quality of service (QoS) requirements for distributed applications deployed in heterogeneous, dynamically changing environments spanning multiple administrative domains.

  An attractive solution is to create an infrastructure which satisfies user QoS requirements by automatically and transparently adapting distributed applications to any environment changes with minimum user input. However, successful use of this approach requires overcoming three challenges: (1) Capturing the application behavior and its relationship with the environment as a set of compact local specifications, using both general, quantitative (e.g., CPU usage) and qualitative (e.g., security) properties. Such information should be sufficient to reason about the global behavior of the application deployment. (2) Finding the ``best'' application deployment that satisfies both application and user requirements, and the various domain policies. The search algorithm should be complete, efficient, scalable with regard to application and network sizes, and guarantee optimality (e.g., resources consumed by applications). (3) Ensuring that the found deployments are practical and efficient, i.e., that the efficiency of automatic deployments is comparable with the efficiency of hand-tuned solutions.

  This dissertation describes three techniques that address these challenges in the context of component-based applications. The modularity and reusability of the latter enable automatic deployments while supporting reasoning about the global connectivity based on the local information exposed by each component. The first technique extends the basic component-based application model with information about conditions and effects of component deployments and linkages, together with interactions between components and the network. The second technique uses AI planning to build an efficient and scalable algorithm which exploits the expressivity of the application model to find an application deployment that satisfies user QoS and application requirements. The last technique ensures that application deployments are both practical and efficient, by leveraging language and run-time system support to automatically customize components, as appropriate for the desired security and data consistency guarantees. These techniques are implemented as integral parts of the Partitionable Services Framework (PSF), a Java-based framework which flexibly assembles component-based applications to suit the properties of their environment. PSF facilitates on-demand, transparent migration and replication of application components at locations closer to clients, while retaining the illusion of a monolithic application.

  The benefits of PSF are evaluated by deploying representative component-based applications in an environment simulating fast and secure domains connected by slow and insecure links. Analysis of the programming and the deployment processes shows that: (1) the code modifications required by PSF are minimal,(2) PSF appropriately adapts the deployments based on the network state and user QoS requirements, (3) the run-time deployment overheads incurred by PSF are negligible compared to the application lifetime, and (4) the efficiency of PSF-deployed applications matches that of hand-crafted solutions.

• Ph.D. Thesis 2004 VALIS: A Multi-language System for Rapid Prototyping in Computational Biology Paxia, Salvatore Abstract | PDF

  Title: VALIS: A Multi-language System for Rapid Prototyping in Computational Biology

  Candidate: Paxia, Salvatore

  Advisor(s): Mishra, Bud

  Abstract:

  Bioinformatics is a challenging area for computer science, since the underlying computational formalisms span database systems, numerical methods, geometric modeling and visualization, imaging and image analysis, combinatorial algorithms, data analysis and mining, statistical approaches, and reasoning under uncertainty.

  This thesis describes the Valis environment for rapid application prototyping in bioinformatics. The core components of the Valis system are the underlying database structure and the algorithmic development platform.

  This thesis presents a novel set of data structures that has marked advantages when dealing with unstructured and unbounded data that are common in scientific fields and bioinformatics.

  Bioinformatics problems rarely have a one-language, one-platform solution. The Valis environment allows seamless integration between scripts written in different programming languages and includes tools to rapidly prototype graphical user interfaces.

  To date the speed of computation of most whole genome analysis tools have stood in the way of developing fast interactive programs that may be used as exploratory tools. This thesis presents the basic algorithms and widgets that permit rapid prototyping of whole genomic scale real-time applications within Valis.

• Ph.D. Thesis 2004 Thick Surfaces: Interactive Modeling of Topologically Complex Geometric Details Peng, Jianbo Abstract | PDF

  Title: Thick Surfaces: Interactive Modeling of Topologically Complex Geometric Details

  Candidate: Peng, Jianbo

  Advisor(s): Zorin, Denis

  Abstract:

  Lots of objects in computer graphics applications are represented by surfaces. It works very well for objects of simple topology, but can get prohibitively expensive for objects with complex small-scale geometrical details.

  Volumetric textures aligned with a surface can be used to add topologically complex geometric details to an object, while retaining an underlying simple surface structure. The simple surface structure provides great controllability on the overall shape of the model, and volumetric textures handle geometric details and topological changes efficiently.

  Adding a volumetric texture to a surface requires more than a conventional twodimensional parameterization: a part of the space surrounding the surface has to be parameterized. Another problem with using volumetric textures for adding geometric detail is the difficulty of the rendering of implicitly represented surfaces, especially when they are changed interactively.

  We introduce thick surfaces to represent objects with topologically complex geometric details. A thick surface consists of three components. First, a base mesh of simple structure is used to approximate the overall shape of the object. Second, a layer of space along the base mesh is parameterized. We define the layer of space as a shell, which covers the geometric details of the object. Third, volumetric textures of geometric details are mapped into the shell. The object is represented as the implicit surface encoded by the volumetric textures. Places without volumetric textures are filled with patches of the base mesh.

  We present algorithms for constructing a shell around a surface and rendering a volumetric-textured surface. Mipmap technique for volumetric textures is explored as well. The gradient field of a generalized distance function is used to construct a non-self-intersecting shell, which has other properties desirable for volumetric texture mapping. The rendering algorithm is designed and implemented on NVIDIA GeForceFX video chips. Finally we demonstrate a number of interactive operations that these algorithms enable.

• Ph.D. Thesis 2004 TM-LPSAT: Encoding Temporal Metric Planning in Continuous Time Shin, Ji-Ae Abstract | PDF

  Title: TM-LPSAT: Encoding Temporal Metric Planning in Continuous Time

  Candidate: Shin, Ji-Ae

  Advisor(s): Davis, Ernest

  Abstract:

  In any domain with change, the dimension of time is inherently involved. Whether the domain should be modeled in discrete time or continuous time depends on aspects of the domain to be modeled. Many complex real-world domains involve continuous time, resources, metric quantities and concurrent actions. Planning in such domains must necessarily go beyond simple discrete models of time and change.

  In this thesis, we show how the SAT-based planning framework can be extended to generate plans of concurrent asynchronous actions that may depend on or make change piecewise linear metric constraints in continuous time.

  In the SAT-based planning framework, a planning problem is formulated as a satisfiability problem of a set of propositional constraints (axioms) such that any model of the axioms corresponds to a valid plan. There are two parameters to a SAT-based planning system: an encoding scheme for representing plans of bounded length and a propositional SAT solver to search for a model. The LPSAT architecture is composed of a SAT solver integrated with a linear arithmetic constraint solver in order to deal with metric aspects of domains.

  We present encoding schemes for temporal models of continuous time defined in PDDL+: ( i ) Durative actions with discrete and/or continuous changes; (ii) Real-time temporal model with exogenous events and autonomous processes capturing continuous changes. The encoding represents, in a CNF formula over arithmetic constraints and propositional fluents, time-stamped parallel plans possibly with concurrent continuous and/or discrete changes. In addition, we present encoding schemes for multi-capacity resources, partitioned interval resources, and metric quantities which are represented as intervals. An interval type can be used as a parameter to action as well as a fluent type.

  Based on the LPSAT engine, the TM-LPSAT temporal metric planner has been implemented: Given a PDDL+ representation of a planning problem, the compiler of TM-LPSAT translates it in a CNF formula, which is fed into the LPSAT engine to find a solution corresponding to a plan for the planning problem. We also have experimented on our temporal metric encodings with other decision procedure, MathSAT, which deals with propositional combinations of linear constraints and Boolean variables. The results show that in terms of searching time the SAT-based approach to temporal metric planning can be comparable to other planning approaches and there is plenty of room to push further the limits of the SAT-based approach.

• Ph.D. Thesis 2004 Unsupervised Discovery of Extraction Patterns for InformationExtraction Sudo, Kiyoshi Abstract | PDF

  Title: Unsupervised Discovery of Extraction Patterns for InformationExtraction

  Candidate: Sudo, Kiyoshi

  Advisor(s): Grishman, Ralph; Sekine, Satoshi

  Abstract:

  The task of Information Extraction (IE) is to find specific types of information in natural language text. In particular, *event extraction* identifies instances of a particular type of event or fact (a particular "scenario"), including the entities involved, and fills a database which has been pre-defined for the scenario. As the number of documents available on-line has multiplied, entity extraction has grown in importance for various applications, including tracking terrorist activities from newswire sources and building a database of job postings from the Web, to name a few.

  Linguistic contexts, such as predicate-argument relationships, have been widely used as *extraction patterns* to identify the items to be extracted from the text. The cost of creating extraction patterns for each scenario has been a bottleneck limiting the portability of information extraction systems to different scenarios, although there has been some research on semi-supervised pattern discovery procedures to reduce this cost. The challenge is to develop a fully automatic method for identifying extraction patterns for a scenario specified by the user.

  This dissertation presents a novel approach for the unsupervised discovery of extraction patterns for event extraction from raw text. First, we present a framework that allows the user to have a self-customizing information extraction system for his/her query: the Query-Driven Information Extraction (QDIE) framework. The input to the QDIE framework is the user's query: either a set of keywords or a narrative description of the event extraction task.

  Second, we assess the improvement in extraction pattern models. By considering the shortcomings of the prior work based on predicate-argument models and their extensions, we propose a novel extraction pattern model that is based on arbitrary subtrees of dependency trees.

  Third, we address the issue of portability across languages. As a case study of the QDIE framework, we implemented a pre-CODIE system, a Cross-Lingual On-Demand Information Extraction system requiring minimal human intervention, which incorporates the QDIE framework as a component for pattern discovery. In addition, we assess the role of machine translation in cross-lingual information extraction by comparing translation-based implementations.

• Ph.D. Thesis 2004 An Efficient and High-Order Accurate Boundary Integral Solver for the Stokes Equations in Three Dimensional Complex Geometries Ying, Lexing Abstract | PDF

  Title: An Efficient and High-Order Accurate Boundary Integral Solver for the Stokes Equations in Three Dimensional Complex Geometries

  Candidate: Ying, Lexing

  Advisor(s): Zorin, Denis

  Abstract:

  This dissertation presents an efficient and high-order boundary integral solver for the Stokes equations in complex 3D geometries. The targeted applications of this solver are the flow problems in domains involving moving boundaries. In such problems, traditional finite element methods involving 3D unstructured mesh generation expe- rience difficulties. Our solver uses the indirect boundary integral formulation and discretizes the equation using the Nyström method.

  Although our solver is designed for the Stokes equations, we show that it can be generalized to other constant coefficient elliptic partial differential equations (PDEs) with non-oscillatory kernels.

  First, we present a new geometric representation of the domain boundary. This scheme takes quadrilateral control meshes with arbitrary geometry and topology as input, and produces smooth surfaces approximating the control meshes. Our surfaces are parameterized over several overlapping charts through explicit nonsingular C ^∞ parameterizations, depend linearly on the control points, have fixed-size local support for basis functions, and have good visual quality.

  Second, we describe a kernel independent fast multipole method (FMM) and its parallel implementation. The main feature of our algorithm is that it is based only on kernel evaluation and does not require the multipole expansions of the underlying kernel. We have tested our method on kernels from a wide range of elliptic PDEs. Our numerical results indicate that our method is efficient and accurate. Other ad- vantages include the simplicity of the implementation and its immediate extension to other elliptic PDE kernels. We also present an MPI based parallel implementation which scales well up to thousands of processors.

  Third, we present a framework to evaluate the singular integrals in our solver. A singular integral is decomposed into a smooth far field part and a local part that contains the singularity. The smooth part of the integral is integrated using the trape- zoidal rule over overlapping charts, and the singular part is integrated in the polar coordinates which removes or decreases the order of singularity. We also describe a novel algorithm to integrate the nearly singular integrals coming from the evaluation at points close to the boundary.

• Ph.D. Thesis 2004 High Performance Data Mining in Time Series: Techniques and Case Studies Zhu, Yunyue Abstract | PDF

  Title: High Performance Data Mining in Time Series: Techniques and Case Studies

  Candidate: Zhu, Yunyue

  Advisor(s): Shasha, Dennis

  Abstract:

  Note: A significantly improved and expanded description of this material is available in the book High Performance Discovery in Time Series Springer Verlag 2004 ISBN 0-387-00857-8.

  As extremely large time series data sets grow more prevalent in a wide variety of settings, we face the significant challenge of developing efficient analysis methods. This dissertation addresses the problem in designing fast, scalable algorithms for the analysis of time series.

  The first part of this dissertation describes the framework for high performance time series data mining based on important primitives. Data reduction trasform such as the Discrete Fourier Transform, the Discrete Wavelet Transform, Singular Value Decomposition and Random Projection, can reduce the size of the data without substantial loss of information, therefore provides a synopsis of the data. Indexing methods organize data so that the time series data can be retrieved efficiently. Transformation on time series, such as shifting, scaling, time shifting, time scaling and dynamic time warping, facilitates the discovery of flexible patterns from time series.

  The second part of this dissertation integrates the above primitives into useful applications ranging from music to physics to finance to medicine.

  StatStream
  StatStream is a system based on fast algorithms for finding the most highly correlated pairs of time series from among thousands of time series streams and doing so in a moving window fashion. It can be used to find correlations in time series in finance and in scientific applications.

  HumFinder
  Most people hum rather poorly. Nevertheless, somehow people have some idea what we are humming when we hum. The goal of the query by humming program, HumFinder, is to make a computer do what a person can do. Using pitch translation, time dilation, and dynamic time warping, one can match an inaccurate hum to a melody remarkably accurately.

  OmniBurst
  Burst detection is the activity of finding abnormal aggregates in data streams. Our software, OmniBurst, can detect bursts of varying durations. Our example applications are monitoring gamma rays and stock market price volatility. The software makes use of a shifted wavelet structure to create a linear time filter that can guarantee that no bursts will be missed at the same time that it guarantees (under a reasonable statistical model) that the filter eliminates nearly all false positives.

• Ph.D. Thesis 2003 Comparing and Improving Centralized and Distributed Techniques for Coordinating Massively Parallel Shared-Memory Systems Freudenthal, Eric Abstract | PDF

  Title: Comparing and Improving Centralized and Distributed Techniques for Coordinating Massively Parallel Shared-Memory Systems

  Candidate: Freudenthal, Eric

  Advisor(s): Gottlieb, Allan

  Abstract:

  Two complementary approaches have been proposed to achieve high performance inter-process coordination on highly parallel shared-memory systems. Gottlieb et. al. introduced the technique of combining concurrent memory references, thereby reducing hot spot contention and enabling the bottleneck-free execution of algorithms referencing a small number of shared variables. Mellor- Crummey and Scott introduced an alternative distributed local-spin technique that minimizes hot spot contention by not polling hotspot variables and exploiting the availability of processor-local shared memory. My principal contributions are a comparison of these two approaches, and significant improvements to the former.

  The NYU Ultra3 prototype is the only system built that implements memory reference combining. My research utilizes micro-benchmark simulation studies of massively parallel Ultra3 systems executing coordination algorithms. This investigation detects problems in the Ultra3 design that result in higher-than-expected memory latency for reference patterns typical of busy-wait polling. This causes centralized coordination algorithms to perform poorly. Several architectural enhancements are described that significantly reduce the latency of these access patterns, thereby improving the performance of the centralized algorithms.

  I investigate existing centralized algorithms for readers-writers and barrier coordination, all of which require fetch-and-add, and discovered variants that require fewer memory accesses (and hence have shorter latency). In addition,my evaluation includes novel algorithms that require only a restricted form of fetch-and-add.

  Coordination latency of these algorithms executed on the enhanced combining architecture is compared to the latency of the distributed local-spin alternatives. These comparisons indicate that the distributed local-spin dissemination barrier, which generates no hot spot tra c, has latency slightly inferior to the best centralized algorithms investigated. However, for the less structured readers-writers problem, the centralized algorithms significantly outperform the distributed local-spin algorithm.

• Ph.D. Thesis 2003 Infrastructure Support for Accessing Network Services in Dynamic Network Environments Fu, Xiaodong Abstract | PDF

  Title: Infrastructure Support for Accessing Network Services in Dynamic Network Environments

  Candidate: Fu, Xiaodong

  Advisor(s): Karamcheti, Vijay

  Abstract:

  Despite increases in network bandwidth, accessing network services across a wide area network still remains a challenging task. The difficulty mainly comes from the heterogeneous and constantly changing network environment, which usually causes undesirable user experience for network-oblivious applications.

  A promising approach to address this is to provide network awareness in communication paths. While several such path-based infrastructures have been proposed, the network awareness provided by them is rather limited. Many challenging problems remain, in particular: (1) how to automatically create effective network paths whose performance is optimized for encountered network conditions, (2) how to dynamically reconfigure such paths when network conditions change, and (3) how to manage and distribute network resources among different paths and between different network regions. Furthermore, there is poor understanding of the benefits of using the path-based approach over other alternatives.

  This dissertation describes solutions for these problems, built into a programmable network infrastructure called Composable Adaptive Network Services (CANS). The CANS infrastructure provides applications with network-aware communication paths that are automatically created and dynamically modified. CANS highlights four key mechanisms: (1) a high-level integrated type-based specification of components and network resources; (2) automatic path creation strategies; (3) system support for low-overhead path reconfiguration; and (4) distributed strategies for managing and allocating network resources.

  We evaluate these mechanisms using experiments with typical applications running in the CANS infrastructure, and extensive simulation on a large scale network topology to compare with other alternatives. Experimental results validate the effectiveness of our approach, verifying that (1) the path-based approach provides the best and the most robust performance under a wide range of network configurations as compared to end-point or proxy-based alternatives; (2) automatic generation of network-aware paths is feasible and provides considerable performance advantages, requiring only minimal input from applications; (3) path reconfiguration strategies ensure continuous adaptation and provide desirable adaptation behaviors by using automatically generated paths; (4) both run-time overhead and reconfiguration time of CANS paths are negligible for most applications; (5) the resource management and allocation strategies allow effective setting up shared resource pools in the network and sharing resources among paths.

• Ph.D. Thesis 2003 Enriched Content: Concept, Architecture, Implementation, and Applications Hung-Hsien, Chang Abstract | PDF

  Title: Enriched Content: Concept, Architecture, Implementation, and Applications

  Candidate: Hung-Hsien, Chang

  Advisor(s): Perlin, Ken

  Abstract:

  Since the debut of the World Wide Web, Web users have been facing the following problems:

  [Extended Semantics]
  When we read or study a digital document that we wish to explore further, typically, we interrupt our work to start a search. It costs time.

  [Reverse Hyperlink]
  When we visit a web page, we might be curious about what other hyperlinks point to the visited page. These links would most likely be of related interest. Can we get ``real time'' information about what other pages are pointing to this page?

  [Version Control]
  Many of us have been frustrated and even annoyed when the hyperlink that we follow gives us a ``404 not found'' or the retrieved webpage content is entirely different from the one we have bookmarked. Could we also have access to the past versions even if the hyperlink has been removed or the content has been changed?

  [Composition Assistant]
  Writing is not an easy task. We labor to structure a body of text, sort out ideas, find materials, and digest information. We would like an automated service that can associate the content we have produced with other contexts(on the Web) and bring these web contexts to us for reference.

  In this thesis, we provide a unified framework and architecture, named enriched content, to resolve the above problems. We apply the architecture and show how the enriched content can be used in each application. We demonstrate that this method can be a new way of writing add-on functions for various document applications without having to write individual plug-in for each application or re-write each application. We also briefly discuss possible future development.

• Ph.D. Thesis 2003 A framework for optimistic program optimization Pechtchanski, Igor Abstract | PDF

  Title: A framework for optimistic program optimization

  Candidate: Pechtchanski, Igor

  Advisor(s): Goldberg, Benjamin

  Abstract:

  The problem of program optimization is a non-trivial one. Compilers do a fair job, but can't always deliver the best performance. The expressibility of general-purpose languages is limited, not allowing programmers to describe expected run time behavior, for example, and some programs are thus more amenable to optimization than others, depending on what the compiler expects to see. We present a generic framework that allows addressing this problem in two ways: through specifying verifiable source annotations to guide compiler analyses, and through optimistically using some assumptions and analysis results for the subset of the program seen so far. Two novel applications are presented, one for each of the above approaches: a dynamic optimistic interprocedural type analysis algorithm, and a mechanism for specifying immutability assertions. Both applications result in measurable speedups, demonstrating the feasibility of each approach.

• Ph.D. Thesis 2003 Secure and Robust Censorship-Resistant Publishing Systems Waldman, Marc Abstract | PDF

  Title: Secure and Robust Censorship-Resistant Publishing Systems

  Candidate: Waldman, Marc

  Advisor(s): Mazieres, David

  Abstract:

  In many cases, censoring documents on the Internet is a fairly simple task. Almost any published document can be traced back to a specific host, and from there to an individual responsible for the material. Someone wishing to censor this material can use the courts, threats, or other means of intimidation to compel the relevant parties to delete the material or remove the host from the network. Even if these methods prove unsuccessful, various denial of service attacks can be launched against a host to make the document difficult or impossible to retrieve. Unless a host's operator has a strong interest in preserving a particular document, removing it is often the easiest course of action.

  A censorship-resistant publishing system allows an individual to publish a document in such a way that it is difficult, if not impossible, for an adversary to completely remove, or convincingly alter, a published document. One useful technique for ensuring document availability is to replicate the document widely on servers located throughout the world. However, replication alone does not block censorship. Replicas need to be protected from accidental or malicious corruption. In addition, a censorship-resistant publishing system needs to address a number of other important issues, including protecting the publisher's identity while simultaneously preventing storage flooding attacks by anonymous users.

  This dissertation presents the design and implementation of two very different censorship-resistant publishing systems. The first system, Publius, is a web based system that allows an individual to publish, update, delete and retrieve documents in a secure manner. Publius's main contributions include an automatic tamper checking mechanism, a method for updating or deleting anonymously published content and methods for publishing anonymously hyperlinked content. The second system, Tangler, is a peer-to-peer based system whose contributions include a unique publication mechanism and a dynamic self-policing network. The benefits of this new publication mechanism include the automatic replication of previously published content and an incentive to audit the reliability with which servers store content published by other people. In part through these incentives, the self-policing network identifies and ejects servers that exhibit faulty behavior.

• Ph.D. Thesis 2003 A Qualitative Profile-based Approach to Edge Detection Yen, Ting-jen Abstract | PDF

  Title: A Qualitative Profile-based Approach to Edge Detection

  Candidate: Yen, Ting-jen

  Advisor(s): Yap, Chee

  Abstract:

  Edge detection is a fundamental problem of computer vision and has been widely investigated. We propose a new framework for edge detection based on edge profiles.

  Our model, based on one-dimensional qualitative edge profile fitting and edge consistency, will produce one continuous edge from an initial seed point. A "profile" is defined as a finite cross-section of a two-dimensional image along a line segment. "Edge consistency" means that all the profiles on the same edge should be consistent.

  Appropriate evaluation functions are needed for different types of edge profiles, such as step edges, ramp edges, etc. An evaluation function must meet the requirement that it will produce local minima at the positions where edges of a given type occurs in the profile. Instead of subjective thresholding, image noise is measured statistically and used as a systematic way of filtering false edges. We describe our method as "qualitative edge profile fitting" because it is not based on arbitrary thresolding. Once an edge point is localized, it can be extended into an edge by matching compatible profiles. Two profiles are considered compatible as long as their average di erence is within the noise measurement. Another feature of our approach is its subpixel accuracy. The utilization of profiles and noise-induced threshold selection make tasks such as joining broken edges more objective.

  We develop the necessary algorithms and implement them. Different evaluation functions are constructed for different edge models and experimented on different one-dimensional profiles. The edge detector, using these evaluation functions, is then examined using different images and under different noise conditions.

• Ph.D. Thesis 2002 Expert-Driven Validation of Set-Based Data Mining Results Adomavicius, Gediminas Abstract | PDF

  Title: Expert-Driven Validation of Set-Based Data Mining Results

  Candidate: Adomavicius, Gediminas

  Advisor(s): Tuzhilin, Alexander; Davis, Ernest

  Abstract:

  This dissertation addresses the problem of dealing with large numbers of set-based patterns, such as association rules and itemsets, discovered by data mining algorithms. Since many discovered patterns may be spurious, irrelevant, or trivial, one of the main problems is how to validate them, e.g., how to separate the ``good'' rules from the ``bad.'' Many researchers have advocated the explicit involvement of a human expert in the validation process. However, scalability becomes an issue when large numbers of patterns are discovered, since the expert cannot perform the validation on a pattern-by-pattern basis in a reasonable period of time. To address this problem, this dissertation describes a new expert-driven approach to set-based pattern validation.

  The proposed validation approach is based on validation sequences, i.e., we rely on the expert's ability to iteratively apply various validation operators that can validate multiple patterns at a time, thus making the expert-based validation feasible. We identified the class of scalable set predicates called cardinality predicates and demonstrated how these predicates can be effectively used in the validation process, i.e., as a basis for validation operators. We examined various properties of cardinality predicates, including their expressiveness. We also have developed and implemented the set validation language (SVL) that can be used for manual specification of cardinality predicates by a domain expert. In addition, we have proposed and developed a scalable algorithm for set and rule grouping that can be used to generate cardinality predicates automatically.

  The dissertation also explores various theoretical properties of sequences of validation operators and facilitates a better understanding of the validation process. We have also addressed the problem of finding optimal validation sequences and have shown that certain formulations of this problem are NP-complete. In addition, we provided some heuristics for addressing this problem.

  Finally, we have tested our rule validation approach on several real-life applications, including personalization and bioinformatics applications.

• Ph.D. Thesis 2002 Responsive Thinwire Visualization of Large Geographic Datasets Been, Kenneth Abstract | PDF

  Title: Responsive Thinwire Visualization of Large Geographic Datasets

  Candidate: Been, Kenneth

  Advisor(s): Yap, Chee

  Abstract:

  This thesis describes a web-based, responsive, zooming and panning visual- ization system for a full-featured geographic description of the United States. Current web-based map servers provide, from a visualization standpoint, little more than one static image per page, with hyperlinks for navigation; continuous zooming and panning requires locally stored data. Our primary contribution is a multi-threaded, scalable and responsive client-server architecture that responds to user requests as naturally and quickly as possible, regardless of network band- width reliability. This architecture can be generalized for use in other applica- tions, including non-geographic ones. To this we add a scalable and exible user interface for navigation of multi-scale geographic data, with intuitive zooming and panning, pop-up feature labels, and a user controlled tree-hierarchy of windows. We build software tools and algorithms for translating the U.S. Census Bureau's TIGER data into a format designed for speedy database retrieval and network delivery, and for generalizing the data into multiple levels of detail. Because of anomalies in the TIGER data, this processing requires some human intervention.

• Ph.D. Thesis 2002 Representing and Modifying Complex Surfaces Biermann, Henning Abstract | PDF

  Title: Representing and Modifying Complex Surfaces

  Candidate: Biermann, Henning

  Advisor(s): Zorin, Denis

  Abstract:

  The increasing demand for highly detailed geometric models poses new and important problems in computer graphics and geometric modeling. Applications for complex models range from geometric design and scientific simulations to feature movies and video games.

  We focus on the fundamental problem of creating and manipulating complex surface models. We address the problem by designing an efficient and general surface representation, and develop algorithms for efficient modification of surfaces represented in this form. Our surface representation extends existing subdivision-based representations with explicit representation of sharp features and boundaries, which is crucial in many computer-aided design applications.

  We consider two types of surface modifications: boolean operations on solids bounded by surfaces, and surface pasting. Our technique rapidly and robustly computes an approximate result rather than aiming for the precise solution. At the same time, our approach allows one to trade speed for accuracy, and, in most cases, compute the result with any desired accuracy. The second type of editing operations we consider address the problem of transferring geometric features between different objects. Our technique makes it easy to combine geometric data from various sources (e.g. 3D scanning, CAGD model) into a single model.

• Ph.D. Thesis 2002 On computing the Pareto-optimal solution set in a large scale dynamic network Daruwala, Raoul-Sam Abstract | PDF

  Title: On computing the Pareto-optimal solution set in a large scale dynamic network

  Candidate: Daruwala, Raoul-Sam

  Advisor(s): Mishra, Bud

  Abstract:

  Let G=(V,E) be a graph with time-dependent edges where the cost of a path p through the graph is determined by a vector functions F(p)=[f_1(p),f_2(p), \dots, f_n(p)], where f_1,f_2,...,f_n are independent objective functions. Where n>1 there is no clear idea of what a ``best'' solution is, instead we turn to the idea of Pareto-optimality to define the efficiency of a path. Given the set of paths P through the network, a path p' is Pareto-optimal if for every p in P for all the objective functions (f_i(p) >= f_i(p')).

  The problem of planning itineraries on a transportation system involves computing the set of optimal paths through a time-dependent network where the cost of a path is determined by more than one, possibly non-linear and non-additive, cost function. This thesis introduces an algorithmic toolkit for finding the set of Pareto-optimal paths in time-dependent networks in the presence of multiple objective functions.

  Multi-criteria path optimization problems are known to be NP-Hard, however, by exploiting geometric and periodic properties of the dynamic graphs that model transit networks we show that it is possible to compute the Pareto-optimal solutions sets rapidly without using heuristics. We show that we can solve the itinerary problem in the presence of response time constraints for a large scale graph.

• Ph.D. Thesis 2002 Informative Features in Vision and Learning Rudra, Archisman Abstract | PDF

  Title: Informative Features in Vision and Learning

  Candidate: Rudra, Archisman

  Advisor(s): Geiger, Davi

  Abstract:

  We explore the role of features in solving problems in computer vision and learning. Features captures important domain-dependent knowledge and are fundamental in simplifying problems. Our goal is to consider the universal features of the problem concerned, and not just particular algorithms used in its solution. Such an approach reveals only the fundamental difficulties of any problem. For most problems we will face a host of other specialized concerns. Therefore, we consider simplified problems which captures the essence of our approach.

  This thesis consists of two parts. First, we explore means of discovering features. We come up with an information theoretic criterion to identify features which has deep connections to statistical estimation theory. We consider features to be ``nice'' representations of objects. We find that, ideally, a feature space representation of on image is the most concise representation of an image which captures all available information in it. In practice, however, we are satisfied with an approximation to it. Therefore, we explore a few such approximations and explain their connection to the information-theoretic approach. We look at the algorithms which implement these approximation and look at their generalizations in the related field of stereo vision.

  Using features, whether they come from some feature-discovery algorithm or are hand crafted, is usually an ad hoc process which depends on the actual problem, and the exact representation of features. This diversity mostly arises from the multitude of ways features capture information. In the second part of this thesis, we come up with an architecture which lets us use features in a very flexible way, in the context of content-addressable memories. We apply this approach to two radically different domains, face images and English words. We also look at human performance in reconstructing words from fragments, which give us some information about the memory subsystem in human beings.

• Ph.D. Thesis 2001 Knowledge Discovery in Databases for Intrusion Detection, Disease Classification and Beyond Berger, Gideon Abstract | PDF

  Title: Knowledge Discovery in Databases for Intrusion Detection, Disease Classification and Beyond

  Candidate: Berger, Gideon

  Advisor(s): Mishra, Bud

  Abstract:

  As the number of networked computers grows and the amount of sensitive information available on them grows as well there is an increasing need to ensure the security of these systems. The security of computer networks is not a new issue. We have dealt with the need for security for a long time with such measures as passwords and encryption. These will always provide an important initial line of defense. However, given a clever and malicious individual these defenses can often be circumvented. Intrusion detection is therefore needed as another way to protect computer systems. This thesis describes a novel three stage algorithm for building classification models in the presence of non-stationary, temporal, high dimensional data, in general, and for detecting network intrusion detections, in particular. Given a set of training data records the algorithm begins by identifying "interesting'' temporal patterns in this data using a modal logic. This approach is distinguished from other work in this area where frequent patterns are identified. We show that when frequency is replaced by our measure of "interestingness'' the problem of finding temporal patterns is NP-complete. We then offer an efficient heuristic approach that has proven effective in experiments. Having identified interesting patterns, these patterns then become the predictor variables in the construction of a Multivariate Adaptive Regression Splines (MARS) model. This approach will be justified, after surveying other methods for solving the classification problem, by its ability to capture complex nonlinear relationships between the predictor and response variables which is comparable to other heuristic approaches such as neural networks and classification trees, while offering improved computational properties such as rapid convergence and interpret-ability. After considering a variety of approaches to the problems of over-fitting which is inherent when modeling high dimensional data and non-stationarity, we describe our approach to addressing these issues through the use of truncated Stein shrinkage. This approach is motivated by showing the inadmissibility of the maximum likelihood estimator (MLE) in the high dimensional (dimension >= 3) data. We then discuss the application of our approach as participants in the 1999 DARPA Intrusion Detection Evaluation where we were able to exhibit the benefits of our approach. Finally, we suggest another area of research where we believe that our work would meet with similar success, namely, the area of disease classification.

• Ph.D. Thesis 2001 Algorithms for Rendering in Artistic Styles Hertzmann, Aaron Abstract | PDF

  Title: Algorithms for Rendering in Artistic Styles

  Candidate: Hertzmann, Aaron

  Abstract:

  We describe new algorithms and tools for generating paintings, illustrations, and animation on a computer. These algorithms are designed to produce visually appealing and expressive images that look hand-painted or hand-drawn. In many contexts, painting and illustration have many advantages over photorealistic computer graphics, in aspects such as aesthetics, expression, and computational requirements. We explore three general strategies for non-photorealistic rendering:

  First, we describe explicit procedures for placing brush strokes. We begin with a painterly image processing algorithm inspired by painting with real physical media. This method produces images with a much greater subjective impression of looking hand-made than do earlier methods. By adjusting algorithm parameters, a variety of styles can be generated, such as styles inspired by the Impressionists and the Expressionists. This method is then extended to processing video, as demonstrated by painterly animations and an interactive installation. We then present a new style of line art illustration for smooth 3D surfaces. This style is designed to clearly convey surface shape, even for surfaces without predefined material properties or hatching directions.

  Next, we describe a new relaxation-based algorithm, in which we search for the painting that minimizes some energy function. In contrast to the first approach, we ideally only need to specify what we want, not how to directly compute it. The system allows as fine user control as desired: the user may interactively change the painting style, specify variations of style over an image, and/or add specific strokes to the painting.

  Finally, we describe a new framework for processing images by example, called ``image analogies.'' Given an example of a painting or drawing (e.g. scanned from a hand-painted source), we can process new images with some approximation to the style of the painting. In contrast to the first two approaches, this allows us to design styles without requiring an explicit technical definition of the style. The image analogies framework supports many other novel image processing operations.

• Ph.D. Thesis 2001 Region-based Register Allocation for EPIC Architectures Kim, Hansoo Abstract | PDF

  Title: Region-based Register Allocation for EPIC Architectures

  Candidate: Kim, Hansoo

  Advisor(s): Palem, Krishna

  Abstract:

  Instruction-level parallelism(ILP) is a family of processor and compiler design techniques that speed up execution by allowing individual machine operations. Explicitly Parallel Instruction computing (EPIC) processors evolved in an attempt to achieve high levels of ILP without the hardware complexity. In EPIC processors most of the functions to extract ILP are performed by the compiler. To take advantage higher level of ILP of these architectures, the ILP compiler must use aggressive ILP technique. This opportunity for improved performance comes at the price of increased compilation time.

  As the size of the compilation unit is limited, the compilation time can be reduced. But the limited scope of compilation may restrict the scope of optimization. As a result, the compiler may generate less efficient quality of code. Ideally, we want to get smaller compilation time and the same or better execution time as that obtained using the global approach.

  In this thesis, we address the problem of the compilation time and execution performance trade-off in region-based compilation within the context of the key optimization of register allocation . We demonstrate that schemes designed for region-based allocation perform as well as or even better than schemes designed for global based allocation while having smaller compilation time. To achieve this goal, we propose several innovative techniques which form the core of this thesis.

  We show considerable compilation time savings with comparable execution time performance by synthesizing our techniques in a region-based register allocation. We also explore the relation between the performance of the register allocation and the region size and quantify it. Our research shows selecting the right size of region has the important impact to the performance of register allocation. We proposed the concept of restructuring the regions based on register pressure and discussed how we can estimate the register pressure in order to improve compilation time while maintaining the execution time.

• Ph.D. Thesis 2001 Adversarial Reasoning: A Logical Approach for Computer Go Klinger, Tamir Abstract | PDF

  Title: Adversarial Reasoning: A Logical Approach for Computer Go

  Candidate: Klinger, Tamir

  Advisor(s): Davis, Ernest

  Abstract:

  Go is a board game with simple rules but complex strategy requiring ability in almost all aspects of human reasoning. A good Go player must be able to hypothesize moves and analyze their consequences; to judge which areas are relevant to the analysis at hand; to learn from successes and failures; to generalize that knowledge to other ``similar'' situations; and to make inferences from knowledge about a position.

  Unlike computer chess, which has seen steady progress since Shannon's [23] and Turing's [24] original papers on the subject, progress on computer Go remains in relative infancy. In computer chess, minimax search with [IMAGE ] - [IMAGE ] pruning based on a simple evaluation function can beat a beginner handily. No such simple evaluation function is known for Go. To accurately evaluate a Go position requires knowledge of the life and death status of the points on the board. Since the player with the most live points at the end of the game wins, a small mistake in this analysis can be disastrous.

  In this dissertation we describe the design, performance, and underlying logic of a knowledge­based program that solves life and death problems in the game of Go. Our algorithm applies life and death theory coupled with knowledge about which moves are reasonable for each relevant goal in that theory to restrict the search space to a tractable size. Our results show that simple depth-first search armed with a goal theory and heuristic move knowledge yields very positive results on standard life and death test problems - even without sophisticated move ordering heuristics.

  In addition to a description of the program and its internals we present a modal logic useful for describing strategic theories in games and use it to give a life and death theory and to formally state the rules of Go. We also give an axiomatization for this logic using the modal [IMAGE ] ­calculus [15] and prove some basic theorems of the system.

• Ph.D. Thesis 2001 Machine Level Optimizations for High Level Languages Leung, Allen Abstract | PDF

  Title: Machine Level Optimizations for High Level Languages

  Candidate: Leung, Allen

  Advisor(s): Palem, Krishna

  Abstract:

  Two machine instruction level compiler optimization problems are considered in this work.

  The first problem is time-constrained instruction scheduling, i.e., finding optimal schedules for machine code in the presence of time constraints such as release-times and deadlines. These types of time constraints appear naturally in embedded applications, and also as a side effect of many other compiler optimization problems. While the general problem is NP-hard, we have developed a new algorithm which can optimally handle many P-time solvable sub-instances. In fact, we show that almost all previous algorithms in this related area can be seen as an instance of the priority computation scheme that we have developed. Our work extends and unifies many algorithmic results in classical deterministic scheduling theory related to release-times, deadlines and pipeline latencies.

  The second problem that we investigate in this work is scalar optimizations in machine code. We present a new framework that utilizes static single assignment form (SSA) at the level of individual machine instructions. Complementing the framework, we have also developed new SSA construction algorithms which are faster than previous algorithms, and are very simple to implement.

• Ph.D. Thesis 2001 Exact Geometric Computation: Theory and Applications Li, Chen Abstract | PDF

  Title: Exact Geometric Computation: Theory and Applications

  Candidate: Li, Chen

  Abstract:

  Exact Geometric Computation: Theory and Applications Abstract This dissertation explores the theory and applications of Exact Geometric Computation (EGC), a general approach to robust geometric computing. The contributions of this thesis are organized into three parts. A fundamental task in EGC is to support exact comparison of algebraic expressions. This leads to the problem of constructive root bounds for algebraic expressions. Such root bounds determine the worst-case complexity of exact comparisons. In the first part, we present a new constructive root bound which, compared to previous bounds, can give dramatically better performance in many common computations involving divisions and radical roots. We also improve the well-known degree-measure bound by exploiting the sharing of common sub-expressions. In the second part, we discuss the design and implementation of the Core Library, a C++ library which embraces the EGC approach to robust numerical and geometric computation. Our design emphasizes ease of use and facilitates the rapid development of robust geometric applications. It allows non-specialist programmers to add robustness into new or existing applications with little extra effort. A number of efficiency and implementation issues are investigated. Although focused on geometric computation, the EGC techniques and software we developed can be applied to other areas where it is critical to guarantee numerical precision. In the third part, we introduce a new randomized test for the vanishing of multivariate radical expressions. With this test, we develop a probabilistic approach to proving elementary geometry theorems about ruler-and-compass constructions. A probabilistic theorem prover based on this approach has been implemented using the Core Library. We present some empirical data.
• Ph.D. Thesis 2001 An On-Line Handwriting Recognizer with Fisher Matching, Hypotheses Propagation Network and Context Constraint Models Oh, Jong Abstract | PDF

  Title: An On-Line Handwriting Recognizer with Fisher Matching, Hypotheses Propagation Network and Context Constraint Models

  Candidate: Oh, Jong

  Advisor(s): Geiger, Davi

  Abstract:

  We have developed an on-line handwriting recognition system. Our approach integrates local bottom-up constructs with a global top-down measure into a modular recognition engine. The bottom-up process uses local point features for hypothesizing character segmentations and the top-down part performs shape matching for evaluating the segmentations. The shape comparison, called Fisher segmental matching, is based on Fisher's linear discriminant analysis. The component character recognizer of the system uses two kinds of Fisher matching based on different representations and combines the information to form the multiple experts paradigm.

  Along with an efficient ligature modeling, the segmentations and their character recognition scores are integrated into a recognition engine termed Hypotheses Propagation Network (HPN), which runs a variant of topological sort algorithm of graph search. The HPN improves on the conventional Hidden Markov Model and the Viterbi search by using the more robust mean-based scores for word level hypotheses and keeping multiple predecessors during the search.

  We have also studied and implemented a geometric context modeling termed Visual Bigram Modeling that improves the accuracy of the system's performance by taking the geometric constraints into account, in which the component characters in a word can be formed in relation with the neighboring characters. The result is a shape-oriented system, robust with respect to local and temporal features, modular in construction and has a rich range of opportunities for further extensions.

• Ph.D. Thesis 2001 Continuous Model for Salient Shape Selection and Representation Pao, Hsing-Kuo (Kenneth) Abstract | PDF

  Title: Continuous Model for Salient Shape Selection and Representation

  Candidate: Pao, Hsing-Kuo (Kenneth)

  Advisor(s): Geiger, Davi

  Abstract:

  We propose a new framework for shape representation and salient shape selection. The framework is considered as a low- to middle-level vision process. The framework can be applied to various topics, including figure/ground separation, searching of the shape axis, junction detection and illusory figure finding. The model construction is inspired by the Gestalt studies. They suggest that proximity, convexity, similarity, good continuation, closure, symmetry, etc, are useful for figure/ground separation and visual organization construction. First, we quantify those attributes for (completed or partial) shapes by our distributed systems. The shape will be evaluated and represented by those results. In particular, the shape convexity, rather than other shape attributes like the symmetry axis or size which were well-studied before, will be emphasized in our discussion. Our problem is proposed in a continuous manner. For the shape convexity, unlike the conventional mathematical definition, we are aimed at deriving a definition to describe a shape ``more convex'' or ``less convex'' than the other. To search the shape axis, more than a binary information telling a point on or off any axis, a continuous information will be obtained. We distinguish axes with ``stronger'' or ``weaker'' declarations. An Easy and natural scheme of pruning can be applied by such representation. For the junction detection, we do not assume any artificial threshold. Instead, the transition from low-curvature to high-curvature curves or curves with discontinuities will be shown by our representation. The model is based on a variational approach, provided by the minimization of the data fitting error as well as the neighborhood discrepancy. Two models will be proposed, the decay diffusion process and the orientation diffusion process.

• Ph.D. Thesis 2001 Language Support for Program Generation Reasoning, Implementation, and Applications Yang, Zhe Abstract | PDF

  Title: Language Support for Program Generation Reasoning, Implementation, and Applications

  Candidate: Yang, Zhe

  Advisor(s): Danvy, Olivier; Goldberg, Benjamin

  Abstract:

  This dissertation develops programming languages and associated techniques for sound and efficient implementations of algorithms for program generation.

  First, we develop a framework for practical two-level languages. In this framework, we demonstrate that two-level languages are not only a good tool for describing program-generation algorithms, but a good tool for reasoning about them and implementing them as well. We pinpoint several general properties of two-level languages that capture common proof obligations of program-generation algorithms:

  • To prove that the generated program behaves as desired, we use an erasure property to reduce the two-level proof obligation to a simpler one-level proof obligation.
  • To prove that the generated program satisfies certain syntactic constraints, we use a type-preservation property for a refined type system that enforces these constraints.

  In addition, to justify concrete implementations, we use a native embedding of a two-level language into a one-level language.

  We present two-level languages with these properties both for a call-by-name object language and for a call-by-value object language with computational effects, and demonstrate them through two classes of non-trivial applications: one-pass transformations into continuation-passing style and type-directed partial evaluation for call-by-name and for call-by-value.

  Next, to facilitate implementations, we develop several general approaches to programming with type-indexed families of values within the popular Hindley-Milner type system. Type-indexed families provide a form of type dependency, which is employed by many algorithms that generate typed programs, but is absent from mainstream languages. Our approaches are based on type encodings, so that they are type safe. We demonstrate and compare them through a host of examples, including type-directed partial evaluation and printf-style formatting.

  Finally, upon the two-level framework and type-encoding techniques, we recast a joint work with Bernd Grobauer, where we formally derived a suitable self application for type-directed partial evaluation, and achieved automatic compiler generation.

• Ph.D. Thesis 2000 SETL for Internet Data Processing Bacon, David Abstract | PDF

  Title: SETL for Internet Data Processing

  Candidate: Bacon, David

  Advisor(s): Schwartz, Jack

  Abstract:

  Although networks and coordinated processes figure prominently in the kinds of data manipulation found in everything from scientific modeling to large-scale data mining, programmers charged with setting up the requisite software systems frequently find themselves hampered by the inadequacy of available languages. The ``real'' languages such as C++ and Java tend to be low-level, requiring the specification of a great deal of often repetitive detail, whereas the higher-level ``scripting'' languages tend to lack the kinds of structuring facilities that lend themselves to the reliable construction of even modestly large systems.

  The high-level language SETL meets both of these needs. Originally conceived as a language which aimed to bring programming a little closer to the idealized world of mathematics, making it extremely useful in the human-to-human communication of algorithms, SETL has proven itself over the years to be an excellent language for software prototyping, primarily because its conciseness and immediacy lend it well to rapid experimentation. These characteristics, together with its general freedom from machine-oriented restrictions, its value semantics, its comprehension-style constructors for aggregates, its skill with strings, and especially its syntactic support for mappings, also make it well suited to high-level data processing.

  In order to play the role of a full-fledged modern data processing language, however, SETL had to acquire the ability to manipulate processes and communicate with them easily, and furthermore to be able to work with networks, particularly the client-server model that rules the Internet. Accordingly, I have integrated a full set of process and network management features into SETL. In my dissertation, I show how the liberal use of fullweight processes, with the high, protective walls that surround them, sustains a modular design approach which in turn provides a strong defense against the main hazards of distributed computing, namely race conditions and deadlock, while preserving the luxury and convenience of programming in a truly high-level language. To this end, I have evolved protocols and design patterns for developing multiplexing servers and clients in SETL, and in my talk, will present examples of fairly complex systems where hierarchies of processes communicate over the network. Such systems tend to be notorious for their unreliability, but in these instances, robustness seems to follow naturally from the readability of simple programs written in an ancient and friendly language.

• Ph.D. Thesis 2000 A Rigorous Framework for Fully Supporting the IEEE Standard for Floating-Point Arithmetic in High-Level Programming Languages Figueroa, Sam Abstract | PDF

  Title: A Rigorous Framework for Fully Supporting the IEEE Standard for Floating-Point Arithmetic in High-Level Programming Languages

  Candidate: Figueroa, Sam

  Advisor(s): Dewar, Robert

  Abstract:

  Processors conforming to the IEEE Standard for Floating-Point Arithmetic have been commonplace for some years, and now several programming languages seem to support or conform to this standard, from hereon referred to as ``the IEEE Standard.'' For example, The Java Language Specification by Gosling, Joy, and Steele, which defines the Java language, frequently mentions the IEEE Standard. Indeed, Java, as do other languages, supports some of the features of the IEEE Standard, including a couple floating-point data formats, and even requires (in section 4.2.4 ``Floating-Point Operations'' of the aforementioned book) that ``operators on floating-point numbers behave exactly as specified by IEEE 754.''

  Arguing that the support current languages offer is not enough, this thesis establishes clear criteria for what it means to fully support the IEEE Standard in a programming language. Each aspect of the IEEE Standard is examined in detail from the point of view of how various arithmetic engines implement that aspect of the IEEE Standard, how different languages (and implementations thereof) support it, and what the range of options are in supporting that aspect. Practical recommendations are then offered (particularly, but not exclusively, for Ada and Java), taking, for example, programmer convenience and impact on performance into consideration. A detailed model specification following these recommendations is provided for the Ada language.

  In addition, a variety of issues related to the floating-point aspects of programming languages are discussed, so as to serve as a more complete guide to language designers. One such issue is floating-point expression evaluation schemes, and, more specifically, whether bit-for-bit identical results are actually achievable on a variety of platforms that conform to the IEEE Standard, as the Java language promises. Closely tied to this issue is that of double rounding, which occurs when a (possibly intermediate) result is rounded more than once before subsequent use or before being delivered to its final destination. So this thesis discusses when double rounding makes a difference, how it can be avoided, and what the performance impact is in avoiding it.

• Ph.D. Thesis 2000 A Language-Theoretic Approach to Algorithms Goyal, Deepak Abstract | PDF

  Title: A Language-Theoretic Approach to Algorithms

  Candidate: Goyal, Deepak

  Advisor(s): Paige, Bob

  Abstract:

  An effective algorithm design language should be 1) wide-spectrum in nature, i.e. capable of expressing both abstract specifications and low-level implementations, and 2) "computationally transparent", i.e. facilitate accurate estimation of time and space requirements. The conflict between these requirements is exemplified by SETL which is wide-spectrum, but lacks computational transparency because of its reliance on hash-based data structures. The first part of this thesis develops an effective algorithm design language, and the second demonstrates its usefulness for algorithm explanation and discovery.

  In the first part three successively more abstract set-theoretic languages are developed and shown to be computationally transparent. These languages can collectively express both abstract specifications and low-level implementations. We formally define a data structure selection method for these languages using a novel type system. Computational transparency is obtained for the lowest-level language through the type system, and for the higher-level languages by transformation into the next lower level. We show the effectiveness of this method by using it to improve a difficult database query optimization algorithm from expected to worst-case linear time. In addition, a simpler explanation and a shorter proof of correctness are obtained.

  In the second part we show how our data structure selection method can be made an effective third component of a transformational program design methodology whose first two components are finite differencing and dominated convergence. Finite differencing replaces costly repeated computations by cheaper incremental counterparts, and dominated convergence provides a generalized iteration scheme for computing fixed-points. This methodology has led us to a simpler explanation of a complex linear-time model-checking algorithm for the alternation-free modal mu-calculus, and to the discovery of an O ( N ³ ) time algorithm for computing intra-procedural may-alias information that improves over an existing O ( N ⁵ ) time algorithm.

• Ph.D. Thesis 2000 Supporting a Flexible Parallel Programming Model on a Network of Non-Dedicated Workstations Huang, Shih-Chen