CSL Decision and Control Group

Title: Stratified Deep Reinforcement Learning Games

Abstract: In this talk we will present some ongoing work exploring the implicit representations that deep networks develop in their latent spaces. Specifically, we will show how a transformer-based reinforcement-learning (RL) agent tasked with solving simple MiniGrid-like games embeds visual observations not on a smooth manifold but on a stratified embedding space whose local dimensionality varies with sub-strategic execution and environment complexity. We will argue that the ambient space of these games is itself stratified, making a stratified latent representation a natural match. These insights into latent-space structure may, in turn, lead to new understandings of how deep models learn and generalize.

Bio: Alberto Speranzon is an LM Technical Fellow and Chief Scientist for Autonomy at Lockheed Martin's Advanced Technology Labs (ATL). At ATL, his work focuses on embedding neurosymbolic methods into autonomous-system design, emphasizing scalability, compositional reasoning, and assurance. Prior to that, he served as a Technical Fellow at Honeywell Aerospace, leading research on perception, navigation, world-modeling, and decision-making for next-generation aerospace systems. He has held editorial positions with IEEE Transactions on Control Systems Technology and currently serves as an Associate Editor for the IEEE Open Journal of Control Systems (more info at https://www.linkedin.com/in/albertosperanzon/)

Location & Time: CSLB02, April 22, 3-4PM. Reception at 2:30PM outside B02.