Abstract
How does the brain flexibly integrate the multiple sources of information needed to control arbitrary goal-directed behavior? Mixed selectivity theory argues that this cognitive flexibility is achieved through flexible neural representations, with most neurons encoding nonlinear (and in some articulations dynamic) combinations of the stimulus factors. In this view, only fundamental computations underlying many behaviors merit neurons dedicated specifically to them. Despite its importance, the question of how mixed representations shape behavior in an attention-demanding task remains open. Our study applies mixed selectivity theory to visual attention by analyzing three factors known to bias saccade target selection during search: bottom-up feature contrast, top-down target guidance, and the history of previous object fixation (inhibitory tagging). We analyzed how single neuron responses in the rhesus superior colliculus encode these three attention-guiding properties of an object landing in the response field during eye movements in visual search, then determined mixed selectivity using two methods: standard nested GLM and our extension of an application of partial information decomposition (PID) to this behavior. We found that (1) Our application of PID, in contrast to standard GLM analyses, captures the dynamics of neural selectivity over time and the subtleties of how a neuron mixes multiple variables. (2) There is ample evidence for cells that sustain their encoding of multiple factors, and also cells whose selectivity varies over the time course of target selection. (3) In addition to these mixed selectivity neurons, a substantial group of neurons is uniquely selective to whether stimuli were previously fixated while searching, suggesting that inhibitory tagging may be a fundamental computation supporting overt visual attention. We conclude that both static and dynamic forms of mixed selectivity are used to represent attention biases in the superior colliculus, and that the colliculus may participate in a neural circuit dedicated to inhibitory tagging.