We all possess a mental library of schemas that specify how different types of events unfold. How are these schemas acquired? A key challenge is that learning a new schema can catastrophically interfere with (i.e., overwrite) old knowledge. One solution to this dilemma is to use interleaved training to learn a single representation that accommodates all schemas. However, another class of models posits that catastrophic interference can be avoided by splitting off new representations when large prediction errors occur. A key differentiating prediction is that, according to splitting models, catastrophic interference can be prevented even under blocked training curricula. We conducted a series of semi-naturalistic experiments and simulations with Bayesian and neural network models to compare the predictions made by the “splitting” versus “non-splitting” hypotheses of schema learning. We found better performance in blocked compared to interleaved curricula, and explain these results using a Bayesian model that incorporates representational splitting in response to large prediction errors. In a follow-up experiment, we validated the model prediction that inserting blocked training early in learning leads to better learning performance than inserting blocked training later in the learning process. Our results suggest different learning environments (i.e., curricula) play an important role in shaping schema composition. We discuss the different roles prediction errors have for “carving nature at its joints”.

The N-back task is often considered to be a canonical example of a task that relies on working memory (WM), requiring both active maintenance of representations of previously-presented stimuli and also processing of these representations. In particular, the set-size effect in this task (e.g., poorer performance on 3-back than 2-back judgments), as in others, is commonly interpreted as indicating that the task relies on active maintenance in a limited-capacity WM system. Here, we consider the alternative possibility that retention in episodic memory (EM) rather than working memory can account for both set-size and lure effects in the N-back task. Accordingly, performance in the N-back task may reflect engagement of the processing ("working") function of WM but not necessarily the retention ("memory") function of WM. To demonstrate this point using a neural network architecture, we constructed a model that was augmented with an EM component, but lacked any capacity to actively maintain information in WM, and trained it to perform the N-back task. We show that this model can account for the set-size and lure effects obtained in an N-back study by Kane et al. (2007), and that it does so as a result of the well-understood effects of temporal distinctiveness on EM retrieval. These findings add to a growing body of evidence that performance on tasks commonly assumed to rely on WM may, in addition or alternatively, rely on EM.