Abstract
The distributed architecture of the cortex poses a fundamental challenge for reinforcement learning: how to assign credit specifically to regions that contribute to successful behavior? Cortical neurons can be driven by both global reinforcers, like rewards, and local sensory features, making it difficult to disentangle these influences. To address this, we investigated cortical reinforcement learning by manipulating the reward-predictive sensory modality during learning tasks, while monitoring key regulators of cortical activity—local inhibitory neurons, and cholinergic inputs. We found that VIP interneurons are broadly recruited by reward-predictive cues via a modality-independent cholinergic signal. However, when task demands aligned with local computation, SST interneurons suppressed VIP recruitment through an inhibitory feedback loop. A computational model demonstrates that this cholinergic-VIP-SST interneuron circuit motif enables targeted reinforcement learning and region-specific credit assignment in the cortex. These results offer a neurobiologically-grounded framework for how the cortex uses global reinforcement signals to direct plasticity to task-relevant regions, enabling those regions to adapt and fine-tune their responses.
Competing Interest Statement
Balazs Rozsa is a founder of Femtonics Ltd. and a member of its scientific advisory board