A neural-network system for control of eye movements: basic mechanisms.

This paper presents a neural-network-based system that can generate and control movements of the eyes. It was inspired by a number of experimental observations on the saccadic and gaze systems of monkeys and cats. Because of the generality of the approach undertaken, the system can be regarded as a demonstration of how parallel distributed processing principles, namely learning and attractor dynamics, can be integrated with experimental findings, as well as a biologically inspired controller for a dexterous robotic orientation device. The system is composed of three parts: a dynamic motor map, a push-pull circuitry, and a plant. The dynamics of the motor map is generated by a multi-layer network that was trained to compute a bidimensional temporal-spatial transformation. Simulation results indicate (1) that the system is able to reproduce some of the properties observed in the biological system at the neural and movement levels and (2) that the dynamics of the motor map remains stereotyped even when the motor map is subject to abnormal stimulation patterns. The latter result emphasizes the role of the topographic projection that connects the motor map to the push-pull circuitry in determining the features of the resulting movements.