Opposing plasticity mechanisms enhance visual inputs that violate their context

Abstract

To navigate complex environments, the mammalian visual system prioritizes informative inputs by comparing incoming sensory signals with learned contextual information. While this process is often conceptualized as a difference computation, cortical circuitry poses fundamental challenges: visual contextual feedback is predominantly amplifying, and whether inhibitory circuits support selective subtraction is unknown. Here, we show how layer 2/3 pyramidal neurons in mouse primary visual cortex can implement this comparison through plasticity that acts in opposite directions on responses to direct visual stimulation and contextual signals. To dissociate these components, we presented full images to mice or images where one quadrant was occluded, removing direct visual stimulation while preserving surrounding contextual information, and measured pyramidal cell responses to these stimuli before and after visual familiarization. After familiarization to non-occluded images, responses to direct visual inputs were selectively reduced, whereas contextual responses to occluded images strengthened. This strengthening generalized across natural scenes, allowing contextual inputs to also drive responses to novel occluded images and amplify responses to novel non-occluded images. A computational model reproduced these changes by weakening feedforward inputs and strengthening both feedforward-driven inhibition and contextual feedback in response to repeated pyramidal cell activity. This context-contrasting mechanism produced a conserved structure of responses to occluded and non-occluded images across mice and previously acquired monkey and human data. Together, these results suggest a mechanism for saliency assignment that exploits the compartmental organization of cortical pyramidal neurons: they learn the generalized statistical structure of familiar contexts, allowing selective amplification of improbable inputs or omissions.