Neuronal computations shaping visual perception

Sensory cortices integrate the information provided by the sensory organs to generate a meaningful interpretation of the world surrounding us. The transformation of sensory inputs into percepts is an essential prerequisite for cognition. Yet, the link between neuronal activities in sensory cortices and perception remains an outstanding question for the field. To investigate this relationship, we perform calcium imaging in the primary visual cortex (V1) of mice performing Go/NoGo orientation discrimination tasks. During this seminar, I will show that visual information in the V1 of mice performing an orientation discrimination task greatly differs from that of naive mice. In naive mice, V1 encodes for the orientation of the stimulus. In trained mice, V1 provides a probabilistic estimate that the visual stimulus is a Go or a NoGo cue. The probabilities computed by V1 faithfully match the mouse's perceptual decision. V1 uses two unsuspected computational tools to compute these probabilities: generalization and categorization. Generalization is generated in V1 L2/3 during perceptual training by a surround suppression mechanism. This surround suppression stabilizes the stimulus representations in V1 around task specific attractors. Categorization is performed by comparing the task attractors' activities evoked by the visual cue presentation. Finally, I will show that generalization is an important component of perceptual learning. Indeed, the absence of generalization precludes perceptual learning. Conversely, strong generalization is linked to better behavioral expertise. Hence, our results suggest that perception is a probabilistic construct resulting from the activation of specific neuronal ensembles or attractors that acquired a behavioral relevance during perceptual training.