Abstract
Despite numerous studies, the neural basis of approximate Bayesian inference, and, in particular, how prior information impacts decisions, remains unclear. A dominant hypothesis is that prior information is incorporated in decision-making at a late stage of processing, in high-order areas such as OFC, ACC or LIP, right before motor commands are issued. Alternatively, information may be broadcast throughout the brain with top-down influences all the way to sensory areas.
To address this question, we examined brainwide neuropixel recordings collected by the International Brain Lab (IBL). In the IBL task, mice are trained to indicate the location of a visual grating stimulus (left or right). Crucially, the prior probability that the stimulus appears on the left flips between 20% and 80% between blocks of variable length.
We found that mice leverage the prior probability over the block to improve their decision accuracy. In particular, they perform better than chance (using this prior) when the grating contrast is set to zero. As a crude approximation to their computation, we therefore designed a Bayes optimal algorithm for estimating the block probability on a trial by trial basis given the specific set of trials experienced by each animal in each session. We then decoded this Bayes optimal estimate from 361 brain regions in the Allen atlas, using carefully quality-controlled recordings of the activity of over 200 000 putative single neurons.
For each brain region, we used cross-validated Lasso linear decoders. Statistical significance was assessed by comparing our result to a null distribution designed to account for potential spurious correlations between blocks and neural drift or other slow changes.
In both inter-trial and within-trial periods, we observed that the prior is widely represented throughout the mouse brain. Consistent with previous work, it is present in particular in high level cortical areas such as ACC or OFC. However, it is also seen throughout substantial portions of the rest of the brain, including early sensory cortical and subcortical regions, such as primary visual cortex or superior colliculus. Overall, we find that around 24% of regions reflect the prior, both in cortical and subcortical regions.
This widespread representation of the prior argues for a neural model of Bayesian inference involving loops between areas, as opposed to a model in which the prior is incorporated only in decision making areas. This study offers the first brain-wide perspective on prior encoding, underscoring the importance of using large scale recordings on a single standardized task.
