Pick Your Channel: Ultra-Sparse Readouts for Recovering Functional Cell Types

ICLR 2026 Conference Submission18968 Authors

Keywords: neural predictive models, sparse readout, functional cell types, retinal ganglion cells, primary visual cortex

TL;DR: We implement ultra-sparse readouts in deep neural encoding models, enabling them to innately learn functional cell types.

Abstract: Clustering neurons into distinct functional cell types is a prominent approach to understand how the brain integrates information about the external world. In recent years, digitial-twins of the visual system based on deep neural networks (DNNs) have become the de facto standard for predicting neuronal responses to arbitrary stimuli. Such DNNs are designed with a common core that learns a representation of the visual input that is shared across neurons, and a neuron-specific readout that linearly combines the core outputs to predict single neuron responses. Here, we propose a novel way to learn an ultra-sparse readout that, instead of linearly combining the shared core features, learns to pick a single channel for each neuron. For retinal ganglion cells, we find that, unlike the previous unconstrained models, this ultra-sparse readout triggers the neural predictive model to innately learn functional cell types with minimal loss in predictive performance. Furthermore, we show that state-of-the-art adaptive regularization models are unable to find such single channels, and that applying strong regularization to encourage sparse channels not only deteriorates performance but also results in response shrinkage. When applied to primary visual cortex neurons, our model exhibits a larger drop in performance compared to the unconstrained model, perhaps indicating a more continuous organization of neuronal function.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 18968