Reservoir computing with neural activity inputs predicts behavior and neural dynamics in mouse decision-making.

Abstract

A major challenge in neuroscience is to elucidate how the brain processes sensory input to generate behavior, especially given the difficulty of measuring neural activity across the whole brain. To address this limitation, previous studies have used artificial neural networks (ANNs) and modeled decision-making circuits in the brain. Here, we used recorded neural activity from mice as additional inputs to a reservoir computer (RC), together with task-related inputs, enabling the RC to operate in a biologically grounded context. We refer to the RC with this hybrid-input condition as a hybrid-input RC (HRC), and to the RC without neural activity input as an artificial-input RC (ARC). Using spike inputs from cortical or subcortical regions, the HRC predicted body movements of head-fixed mice during a task better than the ARC did. This improvement arose not only from the added mouse neuronal activity, but also from the activity generated within the HRC. The generated activity potentially included unrecorded neuronal activity from mice, which was difficult for the ARC. We suggest that the HRC might generate activity patterns resembling unrecorded neural activity, which may contribute to the prediction of animal movements.