Learning algorithm for saccade model with distributed feedback mechanism

In this paper we propose a new learning rule for a spatiotemporal neural network model of the primate saccadic system with a distributed feedback control mechanism. In our model the superior colliculus is represented as the distributed network model and it provides a dynamic control signal to a lumped brain stem model (Robinson-Gisbergen model). Distributed feedforward and feedback weights between the deeper layer of the superior colliculus model and the brain stem model are trained using a finite time interval learning rule based on a steepest descent method. Simulations are carried out on a 20-cell model for horizontal saccades using eye position feedback and velocity feedback, respectively. The model makes accurate saccades to all target locations over the range 2 to 15 degrees even if disturbance is added to the burst generator in the brain stem model.