Latent knowledge in cortical circuits

Learning is usually inferred from performance, yet behavior need not faithfully reveal what an animal knows. We find that mice acquire sound–reward contingencies within tens of trials, hundreds to thousands of trials before stable discrimination emerges. This temporal gap exposes latent task knowledge that conventional measures of performance miss and allows us to distinguish the neural mechanisms of learning from those governing its expression. Auditory cortex (AC) plays a selective role in this dissociation: it is required for learning but dispensable for expert performance. This role is not explained by sensory re-tuning or map expansion. Instead, learning and performance engage distinct higher-order computations in AC: reward prediction during rapid acquisition and action suppression as behavior stabilizes. These computations are carried by spatially clustered ensembles whose organization is orthogonal to classical stimulus-tuning maps, revealing a cortical architecture structured not only by sensory features but also by the computations required to learn and act. Finally, I will turn from how knowledge is acquired and expressed within a single task to how it is organized and reused as animals learn continually across many tasks. Using our new, fully automated home-cage platform—the Continual Learning Mouse Playground—we find that mice can learn eight distinct tasks sequentially while retaining earlier ones; some generalize immediately to novel combinations of familiar components, a form of compositional inference. These findings provide preliminary evidence that rodents construct and use an internal representation of a multidimensional task space, creating a tractable entry point for investigating how neural circuits discover latent structure across a lifetime of learning. 

Dr. Kishore Kuchibhotla is an Associate Professor at Johns Hopkins University.

NACS Seminars are free and open to the public.