I use RL to improve adaptive perception and foveated perception in embodied agents. Currently, my ideas include developing adaptive sensors that can dynamically switch modalities based upon prediction errors and control systems that can take input from these dynamic sensors. The unifying theme here is that perception, decision making and physical action need to be designed together. I am also interested in deep learning theory, and I write notes on it from time to time.

Adaptive sensing

Sensors that switch modality depending upon what is being sensed using prediction error for a learned "forward" world model.

Foveated perception

Using real-time attention (high resolution) at those portions of a visual scene most likely to matter; this can occur when there has been a distributional shift in the data.

Embodied control

Closed-loop policies are generated and executed in real time based on information collected by the sensors.

World models

ML-trained forward models produce predictions that are compared against observed reality and serve to determine whether or not to modify sensor modalities and/or close the policy loop.

Arbabian LabStanford University
EMERGE LabNew York University
MVA ProgramENS Paris-Saclay
MathematicsSorbonne University