Colloquium Details

Better understanding of algorithmic and statistical matters in generative models

Speaker: Andrej Risteski, MIT

Location: 60 Fifth Avenue 150

Date: March 29, 2019, 11 a.m.

Host: Subhash Khot

Synopsis:

In recent years, one of the areas of machine learning that has seen the most exciting progress is unsupervised learning, namely learning in the absence of labels or annotation. An integral part of the advances have been generative models: probabilistic models capturing a variety of generative processes for high-dimensional data.

With this, accompanying statistical and algorithmic questions have emerged, stemming from all major aspects of generative models: representation (modeling power), learning (fitting a model from raw data) and inference (probabilistic queries and sampling from a known model). Theoretical understanding is increasingly more important, as designing the architecture of the model and tuning training heuristics is getting progressively more difficult, and even diagnosing whether an algorithm has succeeded can be hard.

I will showcase some of my research addressing these questions, in the context of (i) computationally efficient inference using Langevin dynamics in the presence of multimodality; (ii) statistical guarantees for learning distributions using GANs (Generative Adversarial Networks); and (iii) explaining surprising properties of vector representations of words (word embeddings).

Speaker Bio:

Andrej Risteski holds a joint position as the Norbert Wiener Fellow at the Institute for Data Science and Statistics (IDSS) and an Instructor of Applied Mathematics at MIT. Before MIT, he was a PhD student in the Computer Science Department at Princeton University, working under the advisement of Sanjeev Arora. Prior to that he received his B.S.E. degree at Princeton University as well. His work lies in the intersection of machine learning and theoretical computer science. The broad goal of his research is theoretically understanding statistical and algorithmic phenomena and problems arising in modern machine learning.

Notes:

Refreshments will be offered starting 15 minutes prior to the scheduled start of the talk.


How to Subscribe