Policy optimization emerges from noisy representation learning

Jonah Brenner, Chenguang Li, Gabriel Kreiman

Keywords: natural intelligence, reinforcement learning, representation learning, noise

TL;DR: We present a framework for how nervous systems can learn behavioral policies through noisy representation learning.

Abstract: Nervous systems learn representations of the world and policies to act within it. We present a framework that uses reward-dependent noise to facilitate policy optimization in representation learning networks. These networks balance extracting normative features and task-relevant information to solve tasks. Moreover, their representation changes reproduce several experimentally observed shifts in the neural code during task learning. Our framework presents a biologically plausible mechanism for emergent policy optimization amid evidence that representation learning plays a vital role in governing neural dynamics.

Primary Area: reinforcement learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 12431