PhD Thesis Defense: Devdhar Patel, Towards Human-Level Control Through Time-Aware Intelligence
Content
Speaker
Abstract
Artificial intelligence has yet to achieve human-level control. Most existing approaches operate in constrained settings where agents observe, compute, and act at fixed, high frequencies. Scaling such methods often involves increasing the control frequency, which can be computationally expensive and impractical in real-world scenarios. Other approaches, such as supervised imitation learning, rely on costly data collection and often fail to generalize beyond their training environments.
This thesis proposes an alternative path toward intelligent control by drawing inspiration from biological systems and exploring the concept of time-aware intelligence. It begins with a critical review of current
paradigms in control theory, reinforcement learning, and imitation learning, outlining their limitations and highlighting why achieving human-level control remains a challenging problem. To address these challenges, the thesis introduces a novel class of algorithms that incorporate time-awareness through the use of multiple decision-making systems operating at different control frequencies. Using this framework, it presents three brain-inspired algorithms that advance the state of the art in control by improving not only performance, but also efficiency, robustness, and computational cost.
Advisor
Hava Siegelmann
Host
Devdhar Patel