Abstract
Humans adjust their learning rate according to local environmental statistics, however existing models of this process have failed to provide mechanistic links to underlying brain signals. Here, we implement a neural network model that uses latent variables from Bayesian inference to shift a neural context representation that controls the “state” to which feedback is associated. Within this model, behavioral signatures of adaptive learning emerge through temporally selective transitions in active states, which also mimic the evolution of neural patterns in orbitofrontal cortex. Transitions to a previous state after encountering a one-off outlier reduce learning, as observed in humans, and provide a mechanistic interpretation for bidirectional learning signals, such as the p300, that relate to learning differentially according to the source of surprising events. Together, our results demonstrate that dynamic latent state representations can afford normative inference and provide a coherent framework for understanding neural signatures of adaptive learning.
Competing Interest Statement
The authors have declared no competing interest.
