Predictive Remapping in Neural Networks: A Model Based on Corollary Discharge Signals for Visual Continuity Across Saccades

When animals are viewing a stable scene, each saccade changes object locations on the retina, yet animals perceive the scene as stable across saccades. Thus, sensory representations must be updated across saccades to align pre-saccadic and post-saccadic visual inputs. One mechanism by which the brain accomplishes this visual continuity is by the predictive remapping of visual receptive fields: in many visual and oculomotor areas, a subset of neurons predictively responds to visual stimuli that will be in their receptive field by an impending saccade even before the saccade is initiated. Corollary discharge (CD), or efference copy has been shown to originate from the intermediate layer of the superior colliculus, where neurons form an orderly map of saccade direction and amplitude, generate a downward motor command to drive saccades and provide a CD of the saccade command to other brain areas. Moreover, the neurons in the visual cortex have been shown to be tuned to saccade direction. Based on these findings, we propose a model that could perform predictive remapping in a neural network. The inputs to the model network convey information about the impending saccade’s direction and amplitude, consistent with known corollary discharge signals. The network effectively integrates these inputs to shift the receptive field of each neuron, similar to those observed experimentally. Both forward remapping and convergent remapping can be performed by the network depending on the formulation of the network input. The model suggests that during forward remapping, the pairwise relationship of receptive fields between cells should be preserved.