Doctoral Thesis: Representation Learning for Control: Lessons from Partially Observable Linear Dynamical Systems

By: Yi Tian

Thesis Supervisor: Suvrit Sra

Details

• Date: Tuesday, August 13
• Time: 10:00 am - 11:30 am
• Category: Thesis Defense
• Location: 32-G882

Additional Location Details:

Abstract:

In the age of large models, the learning strategy, which determines the scaling exponents of performance versus compute, data, and model parameters in the scaling law, is more critical than ever. An important aspect of the learning strategy is the learning objective, such as next-token prediction for language models. What is its counterpart for learning a world model? Answering this question requires understanding the right objective for learning state representations. Observation-driven, action-driven, and cost-driven objectives have all been explored empirically for learning state representations, but a rigorous understanding of them remains lacking.

In this talk, we will examine the different objectives of learning state representations for control from partial and potentially high-dimensional observations through the lens of linear dynamical systems. In this clean setup, we are able to answer a set of theoretical questions, including whether these learning objectives provably work, with what algorithmic ingredients, under what assumptions, and what they achieve. I will reveal deep connections between these learning objectives and the classical Kalman decomposition known in control theory from the 1960s. Additionally, I will discuss the implications of these results for practitioners, and even for decision-makers in everyday life, e.g., how we should handle uncontrollable or unobservable things in decision-making.

Thesis committee: Suvrit Sra (advisor), Ali Jadbabaie, Russ Tedrake, John Tsitsiklis

Zoom: https://mit.zoom.us/my/yitian

Host

• Yi Tian
• Email: yitian@mit.edu