Organisms gather information from multiple modalities to form flexible and reliable perception of external stimuli of interest. Extensive studies suggest that the brain integrates multisensory signals in an optimal way that is predicted by Bayes’ rule. However, the neural architecture and mechanism underlying this feat is largely unknown. In this thesis, I first explore the properties of Bayesian inference for circular variables. Secondly, I show analytically how a single module neural network performs the unisensory Bayesian inference for flat and non-flat priors. Thirdly, I investigate how multisensory information is encoded in different components of a Bayes-optimal decentralized network architecture. In this architecture, each module is able to function independently and cro...[ Read more ]

Organisms gather information from multiple modalities to form flexible and reliable perception of external stimuli of interest. Extensive studies suggest that the brain integrates multisensory signals in an optimal way that is predicted by Bayes’ rule. However, the neural architecture and mechanism underlying this feat is largely unknown. In this thesis, I first explore the properties of Bayesian inference for circular variables. Secondly, I show analytically how a single module neural network performs the unisensory Bayesian inference for flat and non-flat priors. Thirdly, I investigate how multisensory information is encoded in different components of a Bayes-optimal decentralized 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. A perturbative approach is developed to study the cross-talks in the weakly correlated scenario. The most striking discovery is that the cross-channel feedforward links are antagonistic to the reciprocal links. In general, the reciprocal links are excitatory in the short range but inhibitory in the long range, stabilizing a more concentrated population activity, whereas the cross-channel feedforward links are inhibitory in the short range but excitatory in the long range, reducing the cross-talks from different channels for small disparity and improving integration between channels even when they convey information of moderate disparity. These predictions can be verified in future experiments on the brain. Finally, we optimize the network structure with various types of likelihoods and priors through stochastic gradient descent. The statistical relationship between the difference in the optimal network structures and the difference in the priors and the likelihoods clearly shows that the network can encode multisensory information in a distributed manner. Our results generate testable predictions for future experiments and are likely to be applicable to other artificial systems.