Machine Learning with Infinite Models and Finite Computation

We are undergoing a revolution in data. As computer scientists, we have grown accustomed to constant upheaval in computing resources -- quicker processors, bigger storage and faster networks -- but this century presents the new challenge of almost unlimited access to raw information. Whether from sensor networks, social computing or high-throughput cell biology, we face a deluge of data about our world. We need to parse this information, to understand it, to use it to make better decisions. In this talk, I will discuss my work to confront this challenge, developing new machine learning algorithms that are based on infinitely-large probabilistic graphical models. In principle, these infinite representations allow us to analyze sophisticated and dynamic phenomena in a way that automatically balances simplicity and complexity -- a mathematical Occam's Razor.

Our computers, however, are inevitably finite, so how can we use such tools in practice? I will discuss how my approach leverages ideas from Bayesian statistics to develop practical algorithms for inference in infinite models with finite computation. I will discuss how combining a firm theoretical footing with practical computational concerns gives us tools that are useful both within computer science and beyond.