Ma Xiangyu,
He Wang,
Min Yan,
Wenhao Zhang,
K. Y. Michael WongBayesian Model for Multisensory Integration and Segregation (Link to Google Meeting)
Xiangyu Ma1, He Wang1, Min Yan1, Wen-Hao Zhang2, and K. Y. Michael Wong1
1Hong Kong University of Science and Technology
2University of Pittsburgh
The brain processes information from different sensory modalities in our daily routine, and the neural system should have the ability to distinguish whether different signals originate from the same source. Experimental data suggested that the brain can integrate visual and vestibular cues to infer heading- direction according to Bayesian prediction. In the dorsal medial superior temporal (MSTd) area and the ventral intraparietal (VIP) area, there exist two types of neurons, congruent and opposite neurons. By focusing on a prior distribution of stimuli that is fully correlated, a recent work by Zhang et al. (Zhang, 2019) suggested that those two distinct types of neurons have complementary roles in multisensory integration and segregation. In the proposed distributed network architecture, cues of different modalities are processed by different modules, but the modules are reciprocally connected. Congruent neurons of given preferred stimuli in one module are connected to the congruent neurons in the other module with similar preferred stimuli. In contrast, opposite neurons of given preferred stimuli in the two modules are connected to their counterparts with opposite preferred stimuli. This facilitates the congruent neurons to yield Bayesian posterior estimates of multisensory integration in a broad range of parameters, and the opposite neurons to provide signals dependent on cue disparity, enabling the segregation of cues in subsequent processing. However, in the previous model, there are parameter ranges that the inference can only be approximately Bayesian. Hence, in this work, we will approach the dynamics analytically and propose improvements for achieving more accurate Bayesian inference.
Furthermore, the Bayes-optimality in the previous work was based on a prior distribution of stimuli that is fully correlated, whereas in practice, there are many other scenarios described by priors with more than one components. For example, studies in causal inference consider prior distributions with a correlated and an independent component. In the second part of our work, we propose a neural circuit with additional modules to tackle these cases. In addition, we further illustrate that the network encodes strong evidence for the correlations between the prior information and the network structure. Finally, we discuss how the Bayes factor reveals the potential of our network model as a decision making neural circuit for causal inference.
Reference
Zhang, W.-H. a. (2019, May). Complementary congruent and opposite neurons achieve concurrent multisensory integration and segregation. eLife.