The neural representation of perceptual uncertainty in mouse visual cortex

Perceptual decision making is fraught with uncertainty that comes (at least) in two distinct forms: there can be uncertainty about how a stimulus is perceived, and even for a known stimulus there can be uncertainty about the decision that it licenses. Although putative neural correlates of uncertainty have been described in early sensory areas, these two forms of uncertainty have not been distinguished. Thus, it remains unknown whether and how perceptual uncertainty (presumably computed locally), as opposed to decision uncertainty (presumably computed downstream and conveyed by feedback connections), is represented in sensory cortex. Here, we developed a Bayesian ideal observer model that formalized perceptual and decision uncertainty in a single coherent framework and used it for model-based analyses of behavior and neural responses recorded while mice performed a two-alternative perceptual decision making task. After extensively validating the ideal observer on synthetic data, we fitted it exclusively to animals' behavior during the stimulus period of trials and showed that it accurately predicted their task-related choices in the distinct response period of the same trials. For each trial, the fitted model predicted the level of perceptual and decision certainties that we then used as regressors against simultaneously recorded Ca$^{2+}$ imaging data in the primary visual cortex (V1) during the stimulus period. We characterized both the magnitude and variability of V1 layer 2/3 population responses on single trials. After controlling for a number of directly stimulus- or behavior-related nuisance regressors, using a generalized linear model-approach, we found that perceptual certainty significantly contributed to the variability but not the magnitude of responses, while decision certainty did not significantly contribute to either. These results provide evidence for the representation of perceptual as opposed to decision certainty in V1, and suggest that perceptual uncertainty is encoded by the variability rather than the magnitude of neural responses.