Optimal modular network for multisensory integration

Information from multiple modalities is integrated in the brain in a near-optimal way. Based on a decentralized architecture suggested by physiological and theoretical studies, we investigate how multisensory information is encoded in different components of a Bayes-optimal modular network architecture. In this architecture, each module is able to function independently and cross-talks among them are conveyed by feedforward cross-links and reciprocal links. We found that the unisensory likelihoods are encoded in the same-channel connections and the multisensory prior information is encoded in the cross-talks in a distributed manner. The most striking discovery is that the feedforward cross-links and the reciprocal couplings form an antagonistic pair. The feedforward cross-links are inhibitory in the short range but excitatory in the long range, serving to cancel out noises and improve integration for cues with moderate disparity, whereas the reciprocal links are excitatory in the short range but inhibitory in the long range, stabilizing a more reliable population activity. The complementary role played by different types of cross-talks between multisensory regions can be verified in future experiments on the brain.