S. J. Gershman

The paper shows that dopamine release in the nucleus accumbens during instrumental learning reflects Bayesian inference of probabilistic beliefs about hidden environmental states.

This paper introduces a probabilistic successor representation model that adapts to uncertainty and context switches, capturing animal behavior in contextual tasks and unifying SR theory with hippocampal function.

This paper explains undermatching in decision-making as a result of policy compression to minimize cognitive cost, and links it to dopamine levels in Parkinson's disease.

The paper reanalyzes data and conducts experiments to show that temporal discounting can arise from adaptive mental effort allocation due to simulation noise and rational inattention.

The paper develops a Bayesian model of function learning on graph-structured spaces and shows it better predicts human behavior in reward search tasks than alternative models.

This study identifies perseveration as arising from a trade-off between reward maximization and policy complexity, showing human behavior aligns with theoretical predictions.

The paper proposes a theory explaining systematic deviations in human probabilistic inference by introducing a recognition model that adapts to query distribution.

This study investigates how humans use latent structure in multi-armed bandit tasks to guide exploration and learning, finding that participants leverage functional relationships and develop learning-to-learn strategies.

The paper tests a model of human function learning on graphs by having participants predict passenger volumes in a subway network, finding that diffusion-based generalization outperforms heuristic models.

This paper provides a normative justification for decision by sampling based on efficient coding principles, showing how it accounts for context-dependent decision-making.

This study investigates how humans balance exploration and exploitation by manipulating total and relative uncertainty in a multi-armed bandit task, finding that both random and directed exploration strategies are independently used.

The paper shows humans prefer compositional structures in function learning, finding them more predictable and memorable than non-compositional ones.

The paper shows humans use amortized inference in probabilistic reasoning, reusing summary statistics from previous queries to save cognitive resources.