A recurrent neural network for trans-saccadic spatial updating produces receptive field remapping and suppressed moving hills

Saccade-related remapping of target representations between topographic maps: a neural network study

The goal of this study was to explore how a neural network could solve the updating task associated with the double-saccade paradigm, where two targets are flashed in succession and the subject must make saccades to the remembered locations of both targets. Because of the eye rotation of the saccade to the first target, the remembered retinal position of the second target must be updated if an accurate saccade to that target is to be made. We trained a three-layer, feed-forward neural network to solve this updating task using back-propagation. The network's inputs were the initial retinal position of the second target represented by a hill of activation in a 2D topographic array of units, as well as the initial eye orientation and the motor error of the saccade to the first target, each represented as 3D vectors in brainstem coordinates. The output of the network was the updated retinal position of the second target, also represented in a 2D topographic array of units. The network was trained to perform this updating using the full 3D geometry of eye rotations, and was able to produce the updated second-target position to within a 1 degrees RMS accuracy for a set of test points that included saccades of up to 70 degrees . Emergent properties in the network's hidden layer included sigmoidal receptive fields whose orientations formed distinct clusters, and predictive remapping similar to that seen in brain areas associated with saccade generation. Networks with the larger numbers of hidden-layer units developed two distinct types of units with different transformation properties: units that preferentially performed the linear remapping of vector subtraction, and units that performed the nonlinear elements of remapping that arise from initial eye orientation.