Multi-Time Scale World Models

Explore the concept of Multi-Time Scale World Models in this 45-minute talk by Vaisakh Shaj, presented at Montreal Robotics. Delve into the paper "Multi Time Scale World Models," which was accepted as a Spotlight at NeurIPS 2023. Learn about the challenges intelligent agents face in using internal world models for reasoning and prediction across various time scales. Discover the Multi Time Scale State Space (MTS3) model, a probabilistic formalism designed to learn world models operating at multiple levels of temporal abstractions while handling complex uncertainty predictions. Examine the computational aspects and experimental results, with a focus on action-conditional future predictions spanning several seconds. Gain insights from Vaisakh Shaj, a PhD candidate in Robot Learning at the ALR Lab, Karlsruhe Institute of Technology, whose research centers on probabilistic recurrent state space models and continuous learning in non-stationary environments.

Vaisakh Shaj: Multi-Time Scale World Models