Network Model for Dynamics of Perception with Reservoir Computing and Predictive Coding

The sensory information processing in the brain is achieved by mutual interactions between externally given sensory signals and internally generated neural dynamics rather than by a one-directional bottom-up processing. However, the underlying mechanism of the dynamical properties of the sensory processing largely remains to be explored. Here, we propose a neural network model based on the predictive coding and reservoir computing as a model of the dynamical process of perception. The internal network dynamics of the proposed model is trained so that the network reproduces given multidimensional time courses of sensory signal and the prediction error is sent to higher-order network and is triggering the internal network dynamics. The proposed model may contribute to uncover the mechanism of higher-order cognitive function and can be a basis for the application of the neural dynamics for artificial intelligence.