Predictive abstract task representations support few-shot learning in the rat brain

Abstract

Few-shot learning reveals core mechanisms of flexible cognition and adaptive decision- making in animal behavior by showing how animals rapidly infer rules, categories, or action strategies from sparse experience. Few-shot learning often depends on abstract representations that capture task structure and generalize across similar instances of the same task. The orbitofrontal cortex (OFC) is known to encode abstract representations, but whether these representations can adapt to new contingencies at the rapid timescale of few-shot learning is unknown. We trained rats on a sequence-learning task with recurring structure but changing contingencies. Rats demonstrated few-shot learning, reaching near-optimal performance within 2-3 rewards after each contingency change, exploiting the tasks recurring structure to rapidly generalize across sequence blocks. We recorded OFC activity using Neuropixels probes during task performance. The firing rate of OFC units in the few seconds preceding each poke encoded the task state and predicted the selected action. Remarkably, this predictive activity also encoded the abstract notion of role: the function of an action under the current contingencies. Role representations in the OFC changed on the same rapid timescale as the behavioral changes. We suggest that few-shot learning in this task is supported by the rapid remapping of roles onto actions.