Conference on Cognitive Computational Neuroscience

We compared three models of predictive learning as competing accounts of the neural representation of predictive knowledge in two memory areas. Each model captures distinct computational principles: the successor representation (SR) reflects discounted cumulative frequency, transition probability (TP) reflects pairwise conditional probabilities, and the Kalman filter (KF) captures weakening of weights among competing predictors of the same outcome. We tested the unique predictions of each model on neural adaptation following predictive learning in two neural areas important for temporal relational memory: anterior-lateral entorhinal cortex (alEC) and middle temporal gyrus (MTG). We found that alEC was only and best accounted for by the SR, while MTG responses were behaviorally dependent, reflecting KF in participants exhibiting KF-like behavior and SR otherwise. We suggest that SR and KF learning principles compete for behavioral control but can be simultaneously implemented in different neural areas, potentially allowing the learner to achieve diverse goals.