Discretization of continuous input spaces in the hippocampal autoencoder
Keywords: NeuroAI, Memory, Sparse autoencoders, Hippocampus
TL;DR: Sparse compression of images leads to spatial tuning of individual, interpretable neurons (i.e., place cells), which form useful high-dimensional representations that discretize and tile the image space.
Abstract: Understanding the encoding mechanisms of hippocampal place cells remains a significant challenge in neuroscience. Although sparse autoencoders have been shown to exhibit place cell-like activity, the underlying processes are not fully understood. In this study, we compare spatial representations learned by dense and sparse autoencoders trained on images of 3D environments and find that only sparse autoencoders with orthonormal activity regularization in latent space produce place cells. We then show that this regularization promotes similar images to map onto the same neurons, acting as a locality-sensitive hash function. Notably, we demonstrate that these neurons are visually interpretable through activity clamping and decoding, suggesting the formation of detailed episodic memories at the single-neuron level. We then introduce a novel metric to quantify how neurons discretize the image space into disjoint receptive fields, revealing that sparse autoencoders tile input spaces with minimal overlap. Furthermore, we observe that whereas dense autoencoders generate population codes resembling visual cortex activity near criticality, sparse autoencoders produce higher-dimensional codes, thus suggesting a similar coding strategy in the hippocampus. Extending our approach to the auditory domain, we also replicate the emergence of "frequency place cells" by training sparse autoencoders on audio snippets sampled from a frequency-varying signal, and show that population representations retain the statistical structure of the sample distribution. Lastly, we demonstrate that reinforcement learning agents can leverage these high-dimensional image representations to solve complex spatial-cognitive tasks, despite their inherent brittleness. Overall, our findings elucidate how sparse input compression in autoencoders can give rise to discrete, interpretable memories, establishing an explicit link between episodic memory formation and spatial representations in the hippocampus.
Primary Area: applications to neuroscience & cognitive science
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: 8166
Loading