A DISTRIBUTED MODEL OF THE SACCADIC SYSTEM - THE EFFECTS OF INTERNAL NOISE

A neural network model for the saccadic control system was proposed re cently. In this model, the superior colliculus (SC) was represented as a two-layered neural network, with the second (motor) layer having ex tensive lateral interconnections. The SC network then provided a distr ibuted dynamic control signal to a lumped model of the brainstem burst generator. In this paper, the saccadic model is modified so that it m ore closely reproduces the behavior measured experimentally in the pri mate saccadic system. The burst generator in the earlier model was rep laced by a modified version that bears a stronger resemblance to the p rimate burst generator. The artificial trigger signal of the earlier w ork was replaced by a more neurophysiologically plausible mechanism, i n which temporal initiation of saccadic eye movements is achieved thro ugh the output of the SC network itself. With the help of a new traini ng algorithm that simultaneously updated all feedforward and feedback connection strengths, the revised model was trained not only to elicit realistic horizontal and oblique simulated saccades, but also to prod uce more realistic activity in the model's motor layer units. Finally, temporal noise was incorporated into our model and further changes we re made so that discharges of the motor layer units had the same amoun t of variability as that recorded in neural discharges in the primate SC. The performance of the model in the presence of the injected noise was analyzed for different saccadic paradigms. In each case, the degr ee of scatter in the simulated eye movements resembled that recorded i n monkey under similar behavioral conditions. Based on our results, we draw some potentially important inferences about the operation of the actual saccadic eye movement control system.