Multisensory integration is a critical task for animals as it enables them to make sense of the world around them and increases their chances of survival. Recent brain research identified congruent neurons that may be used for integrating multisensory information. These neurons respond selectively to stimuli from multiple modalities, suggesting they play a crucial role in determining the congruence between sensory inputs.

To model complex real-world environments using a Bayesian framework, I consider composite priors with both correlated and independent components. The resulting probabilistic model can be formulated in two steps: the first step integrates cues from both modalities using the correlated prior component, and the second step combines the correlated component with an additio...[ Read more ]

Multisensory integration is a critical task for animals as it enables them to make sense of the world around them and increases their chances of survival. Recent brain research identified congruent neurons that may be used for integrating multisensory information. These neurons respond selectively to stimuli from multiple modalities, suggesting they play a crucial role in determining the congruence between sensory inputs.

To model complex real-world environments using a Bayesian framework, I consider composite priors with both correlated and independent components. The resulting probabilistic model can be formulated in two steps: the first step integrates cues from both modalities using the correlated prior component, and the second step combines the correlated component with an additional input from the independent cue to yield the integrated inference.

The corresponding neural network architecture consists of two groups of neurons in each area. The first group is congruently connected with its counterpart in the other area, and the second group receives inputs from the first group as well as a direct cue. Both groups of neurons are useful for downstream neural circuits responsible for causal inference. I propose that collective noise improves the accuracy of Bayesian and satisfies the requirements of multisensory integration properties. For causal inference, communication between different modules can be muted by gates controlled by opposite neurons as disparity increases.

A well-designed neural circuit can perform sampling-based inference, which bridges the gap between self-normalized importance sampling algorithms and neural circuits. The Poisson processes of neuronal responses contain all the necessary information about likelihood and posterior probabilities to calculate the Bayes factor, which is the ratio of the likelihood functions of a common source and independent sources. I propose a hypothesis of the correlation between sampling algorithms and neural networks can be tested experimentally and extended to other psychophysics tasks.