What do neurons in the superior colliculus encode during visual search?

The selection of saccade targets is affected by both bottom-up (saliency) and top-down (e.g., target guidance) attention biases, as well as the history of previously fixated locations (e.g., inhibitory tagging). The neural correlates of these factors have been extensively studied, but key to our work is that each factor is known to drive the firing of superior colliculus (SC) neurons, which themselves drive eye movements. We seek to characterize the contribution of these factors in the activity of SC neurons by making novel use of an information theoretic method called partial information decomposition (PID). PID is particularly useful in characterizing whether a neuron’s firing represents redundant or unique information about each factor, enabling the identification of neurons that specifically code certain attention biases. A rhesus monkey was trained to search for a specific target disk in a grid of disks that either had the same color or a different color from the target disk. We computed each disk’s saliency, its goal-relevance, and whether it was previously fixated. We recorded from 89 cells and found strong evidence for intermediate-layer SC neurons uniquely coding both target guidance and inhibitory tagging, with less clear evidence for bottom-up salience. More neurons selectively code inhibitory tagging (32) compared to target guidance (19) and saliency (3), and the inhibitory-tagging information is generally also coded faster. PID can identify cells that carry multiple signals, and we identified 22 cells that encoded synergistic information from multiple factors. We conclude from this first application of PID to attention control that bottom-up salience plays a smaller role than task (target guidance) and fixation-selection history in the representation of information by intermediate-layer SC neurons, with the importance of these latter two factors being so great as to warrant the dedication of neurons to specifically encode these feature and spatial attention biases.