Search Papers | Poster Sessions | All Posters
Poster B40 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
Adaptive learning using attractor switches in RNN
Qin He1, Daniel N. Scott1, Matthew R. Nassar1, Cristian B. Calderon2 (); 1Brown University, 2Centro Nacional de Inteligencia Artificial (CENIA)
Behaving adaptively requires determining when to create new state-action associations, and when to modify existing ones. Normative probabilistic models can accomplish this, but are computationally demanding and require strong assumptions. Which approximations to normative models brains use to avoid these difficulties remains unclear, however. Drawing inspiration from work showing thalamo-cortico-basal-ganglia loop involvement in adaptive learning, we develop and characterize a neural network model that builds new state-action associations via Hebbian learning between cortex and striatum when surprise (computed as the entropy of neural responses) is elevated. We test our model on a predictive inference task including change-points, and show that it captures statistics of normative models, human behavior, and individual differences. The mechanisms in our model may therefore support state-action representation dynamics in-vivo, and differences in them may account for individual differences in adaptive behavior.
Keywords: adaptive learning recurrent neural network attractor network basal ganglia