Abstract
When we select a specific visual feature as a focus of attention, neural responses in early visual areas to similar stimuli are amplified across the entire visual field. Top-down gain modulations originating from the parietal/prefrontal cortex have been suggested as one mechanism underlying this global effect of feature-based attention. However, neurons in the parietal/prefrontal cortex typically have mixed selectivity to multiple features, so propagating top-down feedback via these neurons may also cause gain modulations in early sensory neurons that are tuned to behaviorally irrelevant features. To test this, we used a recurrent spiking neural network (e.g. Bouchacourt&Buschman,2019) consisting of a sensory layer that has random projections to a second ‘random’ layer. The sensory layer consisted of eight ring attractor sub-networks, in which neurons were topographically arranged by stimulus selectivity in a circular feature space. The ‘random’ layer consisted of neurons that were randomly and reciprocally connected to multiple sensory neurons. The connections to multiple sets of sensory neurons gave rise to linear mixed selectivity to multiple features in random layer neurons. As a result, top-down modulations originating in the random layer will excite/inhibit sensory neurons across many sub-networks. To simulate feature-based attention, we adopted a classical feature-based attention task (Treue & Trujillo, 1999), presenting two stimuli in one sub-network and applying attentional gain to one of them. Consistent with previous fMRI findings, we could decode the attended stimulus from firing rate patterns of the unattended, unstimulated sub-networks (Serences & Boynton, 2007). Importantly, these patterns were different from the stimulus-driven pattern when the same stimulus was presented to the unattended sub-networks. This implies that feedback from neurons with high-dimensional tuning imposes structure on unstimulated neurons that is consistently, but idiosyncratically, related to the attended feature. Our findings highlight previously unrecognized consequences of relaying top-down feedback via neurons with high-dimensional tuning functions.