Colloquium Details
Bayesian Machine Learning for Efficient Optimization of Black-box Functions
Speaker: José Miguel Hernández Lobato, Harvard Unviersity
Location: Warren Weaver Hall 1302
Date: February 26, 2016, 11:30 a.m.
Host: Subhash Khot
Synopsis:
Many optimization problems in engineering require trading off multiple objective functions that can only be evaluated through expensive calls to a black-box. Bayesian optimization (BO) methods can solve these problems efficiently by performing less function evaluations than other alternative approaches. To achieve this, BO methods combine 1) efficient data collection strategies with 2) flexible probabilistic models and 3) accurate methods for approximate inference. In this talk I will describe my most recent work in some of these areas. First, I will present new data collection strategies based on information theory. Unlike previous methods, these strategies can make optimal decisions regarding the independent evaluation of different black-box objectives at different locations. I will illustrate the effectiveness of these methods in several optimization problems, including the design of optimal hardware accelerators for neural networks. In the second part of the talk I will focus on using BO to speed up the discovery of optimal molecules. To efficiently solve this problem we have to perform approximate inference in very complicated models such as deep neural networks. For this purpose, I will present "Black-box alpha", a new method for deterministic inference that can be applied to such complex models with very little effort. Black-box alpha generalizes previous methods for deterministic inference (variational Bayes and expectation propagation) and can interpolate between them by tuning a single parameter. This allows us to achieve excellent results when making probabilistic predictions on molecule data with deep neural networks.
Speaker Bio:
Jose Miguel Hernandez Lobato is a postdoctoral fellow in the Harvard School of Applied Sciences and Engineering since September 2014. His research interests are in Bayesian optimization, scalable methods for approximate inference and flexible probabilistic modeling of data. Jose Miguel's research is driven by applications of machine learning to expensive optimal design problems in engineering. His work at Harvard is currently funded by a grant from the Rafael del Pino Foundation. Before joining Harvard, Jose Miguel was a postdoctoral research associate at the Department of Engineering of Cambridge University (UK) were he worked in a collaboration project with the Indian multinational company Infosys Technologies. During this time Jose Miguel also gave lectures on Bayesian Machine Learning at Charles University in Prague (Czech Republic). From December 2010 to May 2011, Jose Miguel was a teaching assistant at the Computer Science Department in Universidad Autonoma de Madrid (Spain), where he obtained his Ph.D. and M.Phil. in Computer Science in December 2010 and June 2007, respectively. Jose Miguel also obtained a B.Sc. in Computer Science from this institution in June 2004, with a special prize to the best academic record on graduation.
Notes:
In-person attendance only available to those with active NYU ID cards.