Unified CEBRA learns shared neural latents across recordings to map single-neuron contributions to population codes

Avatar
Poster
Voice is AI-generated
Connected to paperThis paper is a preprint and has not been certified by peer review

Unified CEBRA learns shared neural latents across recordings to map single-neuron contributions to population codes

Authors

Benquet, C.; Mirzaeri, H.; Schneider, S.; Mathis, M. W.

Abstract

Neural computation is thought to rely on distributed populations; however, most single recording samples capture only a fraction of them, and rare cell types may contribute only a small proportion of neurons per session. Pooling many recordings could reveal how these populations encode information, but each recording samples different neurons, so their contributions cannot be easily compared directly. Here we show that recordings can be combined into a single model of all recorded neurons by aligning them with a shared label, used at both training and inference, placing every neuron's contribution to the population code on a common scale. This approach, called Unified CEBRA, retained interpretable task information across recordings in the monkey and mouse sensorimotor cortex, rat hippocampus, and mouse visual cortex. Pooling thousands of excitatory and inhibitory neurons from hundreds of sessions from mouse visual cortex, we found that the neurons that contributed most to the model's readout of visual features were not only the most strongly tuned, but also those that responded more reliably and were less coupled to co-recorded neurons. Unified CEBRA, therefore, can unveil coding principles studied at the population scale with mappable features to individual neurons from sparse, multi-session recordings using our new approach.

Follow Us on

0 comments

Add comment