Shapeshifters: Auditory cortical neurons switch from polysemantic to monosemantic under anesthesia

Luke Taylor

Keywords: auditory cortex, anesthesia, mechanstic interpretability, polysemantic, monosemantic, feature visualization, sparse autoencoders

TL;DR: Using ANNs, we show that under anesthesia, auditory cortical neurons shift from responding to multiple inputs (polysemantic) to responding to single inputs (monosemantic) due to decoupled connectivity, resulting in a lower-dimensional neural code.

Abstract: General anesthesia transitions the brain from a conscious to an unconscious state, but how does sensory processing differ between these conditions? To address this question, we trained neural network encoding models to predict the responses of auditory cortical neurons to natural sounds in both awake and anesthetized ferrets. Utilizing mechanistic interpretability methods, such as feature visualization, linearization and sparse autoencoders, we analyzed these networks tuning and connectivity to uncover key differences in sensory processing. We found that anesthesia decouples neural connectivity, shifting neurons from polysemantic (responding to multiple inputs) to monosemantic (responding to a single input), resulting in a lower-dimensional population code. These findings illuminate how anesthesia alters neural connectivity and encoding, offering new insights into the neural mechanisms underlying sensory processing.

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: 10682