NEURAL NETWORK MODELING OF EYE COMPENSATION DURING OFF-VERTICAL-AXIS ROTATION

Compensatory eye motion during off-vertical axis rotation of the head in darkness (OVAR), has been modelled with a neural network. The three layered nework was trained with back-propagation to simulate the estimation of head velocity during OVAR. The network produced good estimates within its training range and predicted the eye velocity versus head velocity characteristics in the monkey. Invariance of the compensation to changes in tilt angle, not fully addressed in previous models, was demonstrated by the network, along with a smooth decline in velocity estimate below a threshold angle. This also agreed well with data from the monkey. Study of the internal units of the network provided insight into the manner in which pattern comparison produced the estimates. In addition, the behavior of the network's units suggests types of behavior that might be looked for through unit recordings in the central nervous system during OVAR. © 1990.