Lena L. Kemmelmeier¹, John T. Serences¹; ¹University of California, San Diego

Anticipating a relevant stimulus leads to spatially selective modulations of neural activity in visual cortex, producing a “baseline shift” even before a stimulus appears. These baseline shifts are often interpreted as increased gain in the neurons selective for the spatial position of the attended location. Here, we explore the alternative hypothesis that baseline shifts arise from top-down feedback targeting inhibitory neurons. Targeting inhibitory neurons could stabilize sensory populations and increase their dynamic range, aligning with observed reductions in neuronal variability at stimulus onset. Related fMRI work also shows that the magnitude of baseline shifts scales with anticipated noise in stimulus displays, suggesting the visual system may prepare by suppressing irrelevant inputs. To investigate, we used biologically inspired continuous-time recurrent neural networks (ctRNNs) containing “sensory” layers with tuning properties similar to those found in early visual cortex and a higher-order, PFC-like control layer. We trained the ctRNNs on a selective attention task requiring discrimination at one of two possible spatial locations, and we examined how feedback from the control layer targeted excitatory and inhibitory sensory units. We also manipulated the signal-to-noise ratio (SNR) of the stimulus inputs, and imposed sparsity constraints on the control layer weight matrix to limit processing capacity (i.e. to constrain 'attentional resources'). Pre-stimulus feedback preferentially targeted inhibitory sensory units, especially at the noncued sensory location. Moreover, lower SNR elicited stronger preparatory feedback, paralleling prior fMRI findings. These effects held across regularization strengths, suggesting that the ctRNNs naturally converged on an inhibition-focused strategy across stimulus noise levels and resource constraints. Overall, anticipatory attention may be a regulatory process in which pre-stimulus inhibition increases the dynamic range of neural responses so that evoked responses related to relevant stimuli stand out against background noise.