A geometric representation unveils learning dynamics in primate neurons

Abstract

Primates can quickly and advantageously adopt complex rule-based behaviors. We studied acquisition of rule-based classification while recording single neurons in the dorsal-anteriorcingulate-cortex (dACC) and the Striatum. Monkeys performed trial-by-trial classification on a rich set of multi-cue patterns, allowing de-novo rule-learning with varying conceptual complexities every few days. To examine neural dynamics during the learning itself, we represent each rule with a spanning set of the space formed by the stimulus features. Because neural preference can be expressed by feature combinations, we can track neural dynamics in geometrical terms in this space, allowing a compact universal description of neural trajectories by observing changes in either vector-magnitude and/or angle-to-rule. We find that a large fraction of cells in both regions follow the behavior during learning. Neurons in the dACC mainly rotate towards the rule, suggesting an increase in selectivity that approximates the rule; whereas in the Putamen we additionally find a prominent magnitude increase, suggesting strengthening of confidence. Moreover, magnitude increases in the striatum followed rotation in the dACC, and finally, the neural policy at the end of the session predicted next-day behavior. Using this novel framework enables tracking of neural dynamics during learning and suggests differential roles of confidence and policy for the different brain regions.