Real-cyber hybrid neural network for predicting neural circuits in mouse decision-making

A major challenge in neuroscience is to elucidate how the brain processes sensory input to generate behavior, especially given the difficulty to measure 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 developed a real-cyber hybrid neural network (HNN) that directly input the activity of real neurons to an ANN, imposing biological constraints on the model. Using spike inputs from cortical or subcortical regions, the HNN predicted body movements of head-fixed mice during a task better than the ANN did. This improvement arose not only from the added mouse neuronal activity, but from the activity generated within the HNN. The generated activity potentially included unrecorded neuronal activity from mice, which was difficult for the ANN. We propose that HNNs develop brain-like activity to predict animal movements and compensate for unrecorded neural activity.