A recurrent neural network that produces predictive spatial updating using retinal error and eye velocity efference copy signals

*G. P. KEITH1,2, G. BLOHM2, J. D. CRAWFORD1,2,3
1Psychology, York University, Toronto, ON, CANADA, 2Centre for Vision Research, Centre for Vision Research, Toronto, ON, CANADA, 3Biology and Kinesiology, York University, Toronto, ON, CANADA.

It is currently believed that remembered visual target locations are stored in eye-centered coordinates and updated across eye movements. Neuronal behavior associated with this updating has been observed in brain areas associated with saccade generation, in the form of transient receptive field remapping prior to and during the saccade. The dynamics of this remapping, however, remain a question; for example, whether these receptive fields spread or jump (Wurtz & Sommer 2005). We trained three 3-layer recurrent neural networks with discrete time-steps to examine how representations of target position evolve during saccade-related updating. Target position during fixations was represented in the output layer as a hill of activation in a 2-D topographic array of units. Network inputs were initial target position, dynamic eye position, and the signal(s) used to drive the updating which, for the three networks were 1) the initial 'cortical' representation of the saccade target, 2) the dynamic 'brainstem' velocity signal of the saccade, and 3) both. In the first network, predictive updating was observed in which the hill of activity jumped directly from initial to remapped target position in a single time-step. In the second network, a gradual shift in the output hill of activation from initial to remapped target position over the duration of the saccade was observed, the hill's amplitude being suppressed during this movement. In the third network, the evolution of the output activation combined that of a jumping and a moving hill. The latency of the remapping for different trials in this network, as measured by the onset of activity at the updated target location, showed a temporal spectrum that spanned the time immediately before and during the saccade, similar to what has been observed neurophysiologically in the frontal eye fields (Umeno and Goldberg 1997). Our model thus shows that the manner in which updating is carried out in a particular brain region depends on the signals used to perform the updating. The use of both initial saccade retinal error and velocity signals to drive updating explains the temporal spectrum of predictive remapping observed in the brain.