Probing expectation signals in sensory decision-making using large-scale two-photon imaging

During decision-making, sensory input is integrated with internal priors to produce appropriate action. Theoretical work has proposed a hierarchical implementation of this process in visual cortex (Rao & Ballard 1999), but it remains unclear whether and how networks of visual cortical neurons might encode priors in a predictive manner. To investigate this, we trained mice on a standardised behavioural platform for visual decision-making (The International Brain Laboratory, 2023): mice turned a wheel to indicate whether a visual grating appeared on the left or the right, with prior probability alternating between 0.2 and 0.8 in blocks of variable length. Once they were proficient, a warning sound 1s before stimulus onset allowed them to predict the upcoming stimulus. Meanwhile, a 2-photon random-access mesoscope was used to simultaneously image calcium activity of ~10,000 genetically defined L2/3 neurons across primary and higher-order visual cortex. Mice leveraged the block prior to bias their choices, and were able to use the warning sound to time their responses. We found reliable tuning of visual cortical neurons to visual stimulus features, and confirmed widespread movement-related signals. In mice that had pronounced behavioural biases, the block prior could be decoded from the pre-stimulus activity of a small subset of cortical neurons. These preliminary results point towards a network of cortical neurons that faithfully represent task-related information. Ongoing work will examine how primary and higher-order regions of visual cortex integrate stimulus-predictive information with sensory evidence to guide decisions.