Overcoming distortion in multidimensional predictive representation

Predicting how our actions will affect future events is essential for effective behavior. However, learning predictive relationships is not trivial in a multidimensional world where numerous causes bring any one event about. Here we examine (1) how these multidimensional dynamics may distort predictive learning, and (2) how inductive biases may mitigate these harmful effects. We developed a theoretical framework for studying this problem using a computational successor features model. Model simulations demonstrate how spurious observations arise in such contexts to compound noise in memory and limit the generalizability of learning. We then provide behavioral evidence in human participants for a semantic inductive bias that constrains these predictive learning dynamics based on the semantic relatedness of causes and outcomes. Together, these results show that prior knowledge can shape multidimensional predictive learning, potentially minimizing severe memory distortions that may arise from complex everyday observations.