Intrinsic Population Dynamics are a Neuronal Substrate for Visual Attention

Perception results from a dynamic interplay between the feedforward processing of sensory stimuli and intrinsic neural activity 1, which is often dismissed as noise 2. To tailor perceptual processes to the organisms current needs on a continuous, moment-to-moment basis, intrinsic dynamics - rather than just being noise - have been suggested to reflect prior expectations 3,4, task demands 5, and attentional focus 6. Here, we identify a novel signature of attentive state in which intrinsic, collective neural activity modulates visuospatial attention by dynamically interacting with sensory input within the superior colliculus (SC), a midbrain hub that integrates bottom-up visual input with top-down signals 7. We show that these intrinsic dynamics organize cell activity into structured, blob-like features that are temporally and topographically localized and that rival sensory-evoked responses in strength. These features emerge as animals learn to engage in the visual detection task, and predict behavioral outcomes on a trial-by-trial basis. Although independent of sensory input or overt behavior, the features can be recruited to enhance visual responses four-fold in the attentive state, generating a "dynamic saliency map" that aligns with reaction times and behavioral outcomes. A computational model indicates that tunable blob-like features can arise from local excitatory-inhibitory interactions within the SC, and enhance sensory-evoked responses in line with observations. Together, our results identify state-dependent intrinsic population activity that interacts with sensory input to give rise to a saliency-map-like representation linked to flexible, goal-directed behavior.