Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster E in Poster Session E: Thursday, August 6, 10:30 am – 12:15 pm, Kimmel Center, Shorin & Rosenthal Rooms

Complementary model-free and model-based motor learning systems unify experimental results across scales and species

Lakshmi Narasimhan Govindarajan¹, Sunny Duan¹, Sol Markman¹, Nikasha Patel¹, Ila R Fiete¹, Laureline Logiaco²; ¹Massachusetts Institute of Technology, ²University of Colorado Anschutz

Presenter: Laureline Logiaco

Mammals acquire flexible motor skills that go well beyond species-typical behavior through interactions among distributed brain circuits, yet the division of labor among these circuits remains unclear. We propose a structured neural network model of motor learning and control comprising modules corresponding to primary motor cortex (M1), cerebellum (CB), and integrative sensorimotor cortex (ISC). The model operates under key constraints of natural motor behavior: continuous control, no supervised motor trajectories, and noisy, delayed, and partially occluded sensory input. We hypothesize that M1 learns a reinforcement-driven controller that is predominantly model-free, whereas ISC and CB learn self-supervised predictive models of sensorimotor dynamics that support state estimation, latent goal estimation, and model-based control. Across challenging tasks, the model acquires robust motor behavior, including rapid responses to perturbations and active sensing when sensory input is unreliable. The systems interact synergistically, especially when sensory observations are sparse: M1 drives exploration and active sensing, thereby facilitating learning in ISC and CB; once ISC and CB acquire predictive control, they can complement M1. The model unifies behavioral, lesion, and neural findings and yields testable predictions about M1–ISC–CB interactions.

Topic Area: Memory, Learning & Knowledge Structures