Figure 4.

Example training experiments. (A) Each column shows a snapshot of activity during a single trial when the SC sheet (lower left corner) has produced its saccadic command (white square in the SC sheet with arrow showing vector of movement) but before the eye has begun to move. Specifically, the snapshots are taken 50 ms after the command and 20 ms before the movement. In this example trial, the eye is initially in the center of the sheet and the target is located, in retinal coordinates, to the right of the center. This is represented in the “Desired Output” sheet, in workspace coordinates, as a target just right of the center. A downward saccadic command would move the retina downward (red arrow) and thereby cause the image of the object to move upward on the Retina sheet (green arrow). The internal representation of eye position (EP) was updated simultaneously with the saccade command (as in Figure 3, “latencies matched” case). The weights from the Retina and SC to the FEF were initialized to be random in the naive network. Through training (moving rightward in the illustrations), the BC sheet gradually reached the Desired Output for this point in time in the trial, after saccade command but before eye movement (and all other points, not shown). During these training iterations, at this point in the trial, the FEF sheet goes through an early phase of activation focused on neurons that represent the approximate retinal location of the target (to the right of center), but then shifts to activation of neurons representing the location where the target will be (upward) after the saccade. In other words, training the system for spatial constancy caused the FEF to remap its visual representation of the target just before the saccade. (B) Training time course of a second network, using the same stimulus location and saccade vector, but different initial random weights in the maps. Shown is the same trial sequence as shown in panel (A), but for brevity, only the middle row. (C) Three more examples using a variety of stimulus locations and saccade vectors. While the fidelity of the final pattern of sheet activity varies between outcomes (sometimes a punctate final representation, sometimes more dispersed), in all cases, the final centroid of neural activation showed presaccadic remapping, in that the Future Field location was better represented than the Receptive Field location at this point in the trial, just before the saccade.