The dynamics of evidence accumulation in expectation-guided perceptual decisions

Prior experiences bias the way we perceive the world by creating expectations, a reference frame for future decisions. Where in the brain these expectations are generated and where they are projected onto the space of possible decisions, it still unknown. In this talk I will present data from rats trained in a two-alternative forced choice auditory task, where the probability to repeat the previous stimulus category was varied between repeating and alternating trial blocks. First, I will show that rats in this task capitalized on the partial predictability of the stimulus sequence by consistently developing a tendency to repeat or alternate their previous response using an internal estimate of the sequence repeating probability. Second, I will describe how these expectations are combined with the integration of the stimulus evidence using the standard Drift Diffusion model (DDM). I will show that the DDM is at odds with the express choices found in the data, i.e. fast responses in which the reaction time does not depend on the stimulus evidence but the choice does. We modified the standard DDM by adding a second urgency integrator and setting that choices are triggered when either the urgency hits the “go bound” or when the stimulus integrator reaches one of the two choice bounds . According to the model rats can decide when to respond before deciding what to response. It predicts therefore that initial choices could be reverted as new information is processed. We thus analyzed the animals’ orienting movements and found that, in a small percentage of trials, their trajectory went towards one side and then reversed the direction of motion to respond on the opposite. These changes of mind occurred mostly in express responses causing an initial choice only based on prior expectations which was then corrected by strong incoming sensory evidence incongruent with that prior. Third. I will show the results from pharmacological inactivation of the prefrontal cortex (PFC) or the dorsomedial striatum (DMS) showing that only inactivation of the DMS diminished significantly the transition bias. Finally, I will show neural recordings in the DMS which demonstrate that neurons dynamically encode the relevant variables that are necessary to build the expectation biases exhibited by our animals.