From Classical Mechanics to Pattern Recognition: Model-Algorithm Co-Design via Efficient Approximations

Author: Li, Albert Hao

Year: 2027

Degree: Dissertation (Ph.D.)

Advisors: Murray, Richard M.; Yue, Yisong

Committee Members: Chung, Soon-Jo; Culbertson, Preston; Yue, Yisong; Murray, Richard M.

Option: Control and Dynamical Systems

DOI: 10.7907/eg1k-ss21

Abstract

Exact descriptions of real systems are rarely computationally tractable, and tractable descriptions are rarely exact. The manner in which one approximates system behavior, computational outcomes, or patterns in data is therefore of primary concern. In turn, the choices made during such approximations dictate which algorithms exist for system analysis, how fast they run, and what guarantees they can carry. The form this design problem takes depends on the origin of the model. When the governing laws are known, an approximation can be constructed deliberately, and the designer chooses which structure it exposes. When only a numerical surrogate is available, its errors are inherited rather than chosen, and the burden shifts to algorithms that can absorb them. When no known governing principles exist at all, the approximation itself must be learned from data, and it is valuable only insofar as it serves the algorithms that consume it. This thesis develops methods exploring all these regimes, beginning from highly-specific challenges motivated by mechanical systems in robotic manipulation and ending with general-purpose models of high-dimensional data like natural videos. Across these settings, recurring themes include computational tractability, robustness to uncertainty, the degree to which physical priors can be leveraged, and the question of how systems and data should be represented. By designing each approximation around the algorithm that consumes it, this thesis achieves state-of-the-art results in a variety of fields ranging from robotic manipulation to video understanding. These include the fastest multi-finger grasp synthesis method of its time, the first real-time sampling-based controllers competitive with reinforcement learning for contact-rich manipulation, and video-text encoders that outperform models trained on significantly more video data.

Files