Equivalent learning emerges from distinct population dynamics across brain networks
Abstract
Learning operates across many brain circuits to equip animals with a diverse repertoire of skills, from complex motor sequences to spatial navigation; at its core, it involves associating population activity patterns with desired outcomes and enabling volitional reactivation of those patterns to drive behavior. Whether these processes rely on shared or circuit-specific dynamical implementations across the brain remains unknown. The key challenge is that different brain regions typically underpin distinct behaviors, making it difficult to disentangle circuit-specific learning mechanisms from region-specific behavioral demands.
Here, we overcome this limitation by using a brain-computer interface (BCI) to make reward delivery contingent on population activity, thereby imposing an identical associative learning problem on two recurrently connected circuits with distinct dynamical regimes: the primary motor cortex (M1) and the hippocampal area CA3. Mice acquired robust volitional control in both regions, and learning yielded shared signatures across circuits: progressive sparsification of population engagement and greater exploration of reward-related activity patterns. The population dynamics underlying these shared outcomes, however, diverged. In M1, activity showed sustained pre-reward excitation, and population trajectories flowed continuously through the reward-associated state; learning progressively reorganized the geometry of population activity to separate behavioral states. In CA3, activity shifted from pre-reward excitation to post-reward inhibition; learning drove the emergence of approach-and-return dynamics converging onto the reward state, with population geometry reducing the separability of distinct behavioral states. Recurrent network models endowed with distinct minimal constraints, chosen to reflect the dominant computational regime associated with each region, captured key features of these shared and region-specific dynamics, indicating that a circuit's computational architecture is sufficient to account for the divergent implementations of learning.
These findings indicate that equivalent learning outcomes can arise from fundamentally different population dynamics reflecting intrinsic circuit constraints rather than conserved neural strategies. This principled degeneracy reveals that learning is not implemented through a single canonical solution, but through multiple circuit-specific dynamical strategies shaped by local circuit architecture.