Neural mechanisms of learning and representing hidden task structures to afford generalization in problem solving – STRUCTGEN

Humans are adept at generalizing previous experiences to find novel solutions to problems. This behavioral flexibility is a hallmark of human intelligence while it is often severely impaired in patients with psychological and mental disorders and remains a key challenge for most artificial intelligence systems. Computational models have suggested that the ability to organize the relationships between entities (e.g., objects, people, or abstract task states) into structural representations in conceptual spaces is critical to afford generalization. Recently, I showed that the hippocampus (HC) and prefrontal cortex (PFC) can construct the “cognitive map” of task structure by organizing piecemeal learned abstract relationships between entities into geometrical representations. However, we know little about how the brain infers a hidden, multidimensional task structure while experiencing unidimensional stream of event sequences. The goal of this project is to elucidate how the HC and PFC interact to build a task structure representation and use it to plan novel decisions. In three work packages, I will address the following questions: (1) Can changes in task representations in the brain account for the neural encoding of the values for model-based decisions? (2) Does generalization ability depend on to what extent one’s subjective task representation reflects the true task structure? (3) Does the brain adopt “replay”, the sequential reactivation of neural activity patterns, to infer the hidden task structure by reconstructing the order of replay sequences? To address this, I will use multimodal neuroimaging (fMRI and MEG) combined with computational models to relate behavior and brain activity of healthy human participants while learning/performing a goal-directed decision making task. By answering these fundamental questions, the findings of this project will contribute to a better understanding of human intelligence.