Advisor: Elad Hazan
Examiners: Sanjeev Arora, Mark Braverman
Readers: Yoram Singer, Karthik Narasimhan
Abstract follows below.
Temporally correlated data and non-convex programs are two core challenges at the frontier of reconciling theory and practice in machine learning. I will present varied perspectives on algorithms and provable learning guarantees in stateful environments, such as those encountered in reinforcement learning, control theory, and time-series analysis. Unified by the framework of online convex optimization, this thesis will consist of two main sections:
- Learning dynamical systems: a line of theoretical work on learning to predict, control, and plan in the presence of time-invariant state transitions, beyond the linear-quadratic-Gaussian assumptions of classical control theory. The regret framework gives us robust algorithms which bypass the usual distributional assumptions and system identification pipelines.
- Optimization in deep learning: a line of empirical work on understanding the behavior of adaptive gradient methods in state-of-the-art deep learning. These poorly-understood heuristics were borne from the online convex world; I will discuss some of the intuitions we can carry over, as well as some of the challenges and insufficiencies.
Finally, I will tie these together with some recent work applying these ideas to my favorite non-convex time series problem: statistical language modeling with deep neural networks.