DYNAMIC UPDATING OF REWARD PREDICTIONS BY THE PARA-HIPPOCAMPAL NETWORK

Goal-directed behavior relies on predictive maps that link stimuli and actions to specific outcomes. We hypothesize that the brain dynamically selects from multiple stimulus-outcome maps to navigate the same physical environment across varying contexts. To test this, we developed a virtual-reality task in which mice repeatedly run through the same environment in which they encounter distinct landmarks along the linear track. Depending on an auditory contextual cue, a reward could be obtained at the first or the second landmark by stopping there. We performed simultaneous Neuropixel recordings in the hippocampus, enthorinal, perirhinal, temporal association, auditory, and visual cortex. A total of 7517 neurons were recorded across these regions in 8 mice. We found that 10-15% of neurons showed anticipatory coding of reward when mice approached landmarks, but only when the landmark was predictive of reward in the current context. Besides single neurons dynamically encoding reward expectancy in their spike rates, we also found consistent temporal sequences of spikes across neurons which did the same thing. Decoding of population activity revealed that contextual information was kept online when the two landmarks were close together in space, but the decoding of context dropped to chance in between the landmarks when they were further apart. In the hippocampus, contextual information ramped up before the animal arrived at the second landmark suggesting that the hippocampus encodes predictive maps of stimulus-outcome associations which are dynamically employed according to the context of the environment.