Towards Stable Learning in Predictive Coding Networks

26 Sept 2024 (modified: 05 Feb 2025)Submitted to ICLR 2025EveryoneRevisionsBibTeXCC BY 4.0
Keywords: Neuroscience, Predictive Coding, Predictive Coding Networks, Neural Networks
TL;DR: This work improves stability and scalability through novel techniques by addressing instability and prediction error concentration in predictive coding networks by analyzing length dynamics.
Abstract:

Predictive coding (PC) offers a biologically plausible model of cortical functions, encompassing processes such as learning, prediction, encoding, and memory. However, predictive coding networks (PCNs) face significant challenges in stability and scalability, which constrain our capacity to elucidate cortical computation. Our study identifies instability in PCNs as a fundamental issue, focusing on the exponential growth of latent state norms and prediction errors after inference. These dynamics lead to exploding and vanishing gradients in PCNs. Moreover, the concentration of prediction errors near the input and output layer impedes effective learning, exacerbating performance degradation as network depth increases. To address these limitations, we propose stabilizing techniques for PCNs, including length regularization and sequential training with skip connection modules. This approach counteracts the exponential growth of latent states and makes the distribution of prediction errors more uniform across layers. Empirical evaluations demonstrate that our approach enhances stability and generalization, enabling the training of deeper networks more efficiently. This study deepens our understanding of complex dynamics in cortical networks, thereby advancing the practical application of predictive coding theory to its full potential.

Supplementary Material: zip
Primary Area: other topics in machine learning (i.e., none of the above)
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: 5667
Loading

OpenReview is a long-term project to advance science through improved peer review with legal nonprofit status. We gratefully acknowledge the support of the OpenReview Sponsors. © 2025 OpenReview