PhD Thesis Defense: Devdhar Patel, Time Aware Intelligence for Efficient and Resilient Control
Content
Speaker
Abstract
Current artificial intelligence (AI) algorithms for control lack temporal awareness: an understanding of time beyond the mere chronology of events. As a result, these systems cannot adapt their outputs to dynamic temporal contexts, which is critical in real-time control where the optimal action may change rapidly with the passage of time. This thesis explores how introducing time-awareness into control algorithms improves exploration, accelerates learning, enhances compute efficiency, and increases robustness to missing data; all while reducing the need for high-frequency decision-making.
I identify three time-awareness mechanisms inspired by the brain: (1) diversity in processing speeds, (2) internal oscillators, and (3) action chunking through fast experience replay. Based on these principles, I propose three novel algorithms that advance the state of the art in continuous control.
Temporally Layered Architecture (TLA) employs multiple policies operating at different step sizes to discover Pareto-optimal tradeoffs between accuracy and energy. Despite training three policies in parallel, TLA converges faster due to improved exploration from slower layers. The resulting policies require significantly fewer decisions and compute while maintaining state-of-the-art performance.
Sensory Layered Architecture (SLA) extends the TLA concept to the sensory input space, using policies with increasing levels of sensory information to discover Pareto-optimal tradeoffs between information and accuracy. The first layer relies solely on internal signals (e.g., previous actions and an oscillator), enabling robust performance even under input occlusion or noise. SLA demonstrates that training with diverse sensory inputs yields emergent robustness not achievable when layers are trained in isolation.
Sequence Reinforcement Learning (SRL) introduces a frequency gap between the actor and critic to learn action chunks—solving a key limitation in continuous control: sensitivity to timestep choice. SRL separates decision and actuation frequencies, dramatically reducing the number of decisions needed without compromising performance. On benchmark tasks, SRL achieves state-of-the-art results at human-level decision rates, paving the way for deployment on low-compute, low-frequency hardware.
Together, these algorithms support efficient and resilient control. TLA and SRL allow for ultra-low decision rates in stable settings and rapid adaptation in unpredictable environments. SLA adds robustness to sensory disruptions and adversarial noise.
Finally, I propose a general framework for building biologically inspired AI under real-world constraints—energy, information, time, damage, and more—as an alternative to the prevailing paradigm of scaling compute, data, and energy. This approach charts a biologically grounded path toward practical, efficient, and adaptive intelligence.
Advisor
Hava Siegelmann
Host
Devdhar Patel