A Mixture of Experts in Associative Generalization

After learning that one stimulus predicts an outcome (e.g., an aqua-colored rectangle leads to shock) and a very similar stimulus predicts no outcome (e.g., a slightly greener rectangle leads to no shock), some participants generalize the predictive relationship on the basis of physical similarity to the predictive stimulus, while others generalize on the basis of the relational difference between the two stimuli (e.g., “higher likelihood of shock for bluer stimuli”). To date, these individual differences in generalization rules have remained unexplored in associative learning. Here, we present evidence that a given individual simultaneously entertains belief in both “similarity” and “relational” rules, and generalizes using a mixture of these strategies. Using a “mixture of experts” modelling framework constrained by participants self-reported rule beliefs, we show that considering multiple rules predicts generalization gradients better than a single rule, and that generalization behavior is better described as switching between, rather than averaging over, different rules.