Neurocomputational modeling of rule abstraction and memorization during probabilistic stimulus-reward learning

Preferential choice among multi-attribute stimuli commonly involves one of the two learning strategies: rule abstraction and memorization. When stimulus features combine in a systematic, albeit complex way to predict rewarding or punishing outcomes (e.g., a combination of color and shape distinguishes edible from poisonous mushrooms), corresponding learning problems can be solved via rule abstraction. In other problems lacking this systematicity, stimuli have to be memorized individually. Here, we use fMRI, eye-tracking, and cognitive modeling to study how humans deploy these two learning strategies to select between two simultaneously presented objects. We observed differential learning trajectories and fixation patterns, indicating sudden rule discovery and incremental learning, respectively, captured by cognitive modeling. The derived process estimates allowed us to identify overlapping brain networks associated with cognitive control and value-based decision-making. Importantly, our multi-modal data and model-informed analyses link those processes to unique brain regions, revealing the neurocognitive mechanisms of rule abstraction and memorization.