Item

Abstract

To select and initiate appropriate actions in complex decision-making tasks requires information to be aggregated and exploited within and across trials. We know little about the brain mechanisms for computation and storage of across-trial information, such as prior probabilities of stimuli or rewarded actions. We examine neural correlates of this information using neural activity recorded in the international brain lab (IBL) task: 583 neuropixel penetrations covering 361 brain regions in the Allen atlas. In the IBL task, mice maximize rewards by exploiting a blockwise prior probability governing the appearance of stimuli. Behavioural modeling has revealed the (usually approximate) way in which the mice are sensitive to the identity of the block. We therefore use a combination of encoding and decoding to localize which brain regions contain either the best estimate of the stimulus prior taken from behavioral modeling or other related quantities such as the true Bayesian posterior over the block.For encoding, we use a Generalized Linear Model (GLM) with a linear link function to express neural activity as a function of numerous task-related variables such as stimulus onset time, contrast, licking, reward delivery, and wheel movement. We incorporate either the block probability or our approximate prior estimate into this model as well, and are examining the variance explained by each regressor.Our decoding approach includes the use of linear decoders, linear discriminant analysis, and a reduced-dimensional representation of the PSTH of neural activity by region to examine how neural activity changes as a function of block probability, stimulus side, or prior estimate and how information is tied together between successive trials.Combining these approaches, we hope to gain insight into the distribution of information about the prior across the brain, as well as clues as to where and how it is computed and updated across trials and how it is brought to bear upon within-trial information processing.Preliminary results indicate substantial variability in the encoding and decodability of prior information across all of: subjects, sessions and brain regions. Despite this, we identify a set of brain regions which still show information content beyond what we expect from said variability. Cortical (retrosplenial) and subcortical (MRN) areas offer some particular promise for investigation, with evidence for distinct sub-populations in the latter integrating prior and stimulus information across different timescales.