From Optimization to Equilibration: A New Framework for AI and ML

08 Mar
Thursday, 03/08/2018 2:00pm to 4:00pm
Computer Science Building, Room 140
Ph.D. Dissertation Proposal Defense
Speaker: Ian Gemp

"From Optimization to Equilibration: A New Framework for AI and ML"

Artificial Intelligence (AI) focuses on the design of agents that act rationally. The Maximum Expected Utility (MEU) principle formalizes the idea of a rational agent as an optimization problem. This principle has pulled optimization to the center of attention in AI and Machine Learning (ML), however, a new paradigm is emerging. Many existing algorithms such as those in Reinforcement Learning (RL) or inference in graphical models can be viewed as solving equilibrium problems. Moreover, some recent ML models, specifically generative adversarial networks (GANs) and its variants, are now explicitly formulated as equilibrium problems.

For this thesis, we propose plans to advance our understanding of equilibrium problems so as to improve state-of-the-art in these and related domains. In particular, we focus on GANs: improving state-of-the-art with heuristics, developing tools for analyzing the dynamics of training, solving a fundamental linear-quadratic variant, and building a framework for online training.

Advisor: Sridhar Mahadevan