On the Identifiability of Latent Action Policies

Sebastien Lachapelle

Track: Extended Abstract Track

Keywords: identifiability, disentanglement, representation learning, latent action policies, imitation learning

TL;DR: We prove an identifiability result for latent action policies under suitable conditions

Abstract: We study the identifiability of _latent action policy learning_ (LAPO), a framework introduced recently to discover representations of actions from video data. We formally describe desiderata for such representations, their statistical benefits and potential sources of unidentifiability. Finally, we prove that an entropy-regularized LAPO objective identifies action representations satisfying our desiderata, under suitable conditions. Our analysis partly explains why _discrete_ action representations are crucial in practice.

Submission Number: 64