Effects of Neural Assembles in Causal Inference Based on an Entropy-Maximization Bayesian Neural Network

Causal inference is an important function of the nervous system. To explore causal inference, Bayesian inference performs as the possible framework, mapping neural implementation onto various cortical areas. Neural assembles participate in Bayesian inference. However, the contributions of neural assemblies in causal inference remain unclear. In this paper, a Bayesian spiking neural network is designed with the entropy-maximization (EM) method to simulate causal inference of visual hidden cues. Hidden cues determine types of visual images in simulations. With images received, the network generates neural spiking trains and modifies its plastic weights with the EM method with constraint conditions. After modifications, the network can identify hidden cues with induced neural responses. Over repeated simulations, similarity and responsivity of neural activities are measured to determine neural assembles. Through principal component analysis of neural responses, contributions of neural assembles in causal inference are explored. During identifications of given stimuli, different neural assemblies make various time-varying contributions. With acceptable performance in causal inference of designed stimuli, the network simulates the emergence of neural assembles and measures their contributions. The Bayesian spiking neural network with the EM method provides the possible framework to explore effects of neural assembles in cause inference.