Stratification Theory for Reinforcement Learning

Abstract: In this talk I will use the framework of poset-stratified spaces to study games, where reward can be both discrete and continuous. Following work by Yuliy Baryshnikov, I will show how certain video games naturally give rise to stratified spaces. Surprisingly, when modern neural nets are trained to play these same video games, a similar stratification structure can be observed in their latent representations. Our methods follow recent work by Michael Robinson and others on using Volume Growth Laws to detect non-manifold structure in the token space for LLMs. We expand and strengthen Robinson’s analysis by considering non-textual data and prove a realization result for volume growth in a stratified space. This is joint work with Brennan Lagasse, Ngoc B. Lam, Gregory Cox, David Rosenbluth, and Alberto Speranzon.

algebraic topologycategory theory

Audience: researchers in the topic

Bilkent Topology Seminar

Series comments: Contact the organizer to get access to Zoom.

Recordings of talks available at www.youtube.com/channel/UCLrmyGpqxyeVpTcA1b5HcMw/videos