Blocked training facilitates learning of multiple schemas

Abstract

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”.

Neural codes track prior events in a narrative and predict subsequent memory for details

“…Our experiment relies on participants' ability to learn the schemas that we presented. We used blocked schema training, as Beukers et al (2023) showed that this is effective in the paradigm that we used. To verify that participants learned the schemas, we tested whether participants learned to distinguish the ritual order in the North schema as opposed to the South schema.…”

“…This sequence design leads to the non-Markovian property that (starting with the transition from stage 2 to stage 3) participants need to represent what happened earlier in the wedding in order to predict what rituals they will observe going forward. In a previous study that measured behavior only (i.e., no fMRI), we demonstrated using these materials that participants could successfully learn the ritual transition structure of the weddings through experience (i.e., without seeing the actual graph shown in Figure 1B) if they were given blocked training (i.e., where they viewed multiple weddings from the North, followed by multiple weddings from the South), but they performed much worse when given interleaved training (i.e., strictly alternating weddings from the North and South; Beukers et al, 2023). Here, we leveraged this blocked training procedure on day 1 of our study to facilitate successful learning of the ritual transition structure.…”

Toward the Emergence of Intelligent Control: Episodic Generalization and Optimization

“…In the remainder of this article we apply the framework to simulate key phenomena observed in three different lines of empirical inquiry that have previously been explained using different formalisms: 1) the effects of reward versus transition revaluation in Momennejad et al's (2017) multi-step learning task, explained in terms of combination of model-based decision making (Daw et al, 2011) and the learning of successor representations (Dayan, 1993;Gershman et al, 2012) in prediction and decision-making; 2) the effects of blocked versus interleaved training in Beukers et al's (2023) state prediction task, explained in terms of latent cause inference (Courville et al, 2006) for event segmentation in the context of continual learning; and 3) the effects of blocked versus interleaved training on the structure of and selection among semantic representations formed in Flesch et al's (2018) categorization task, explained in terms of a variant of Hebbian learning. We show that interactions between context processing and episodic memory are sufficient to reproduce the phenomena in each of these domains, providing a unifying account in terms of a single, integrated set of underlying learning and processing mechanisms.…”

“…This can be overcome by further differentiating the context representations associated with each setting (e.g., learning distinct context representations for restaurants and food courts and using these contexts to guide retrieval of relevant memories). Recent empirical work suggests that people can learn how to do this very effectively, but that this depends on the temporal structure of the environment: people do better when trained in blocks of each situation than when trained interleaved (Beukers et al, 2023).…”

Episodic memory supports the acquisition of structured task representations

“…We developed a task-dependent sequence-learning environment (Fig. 2a) based on the paradigm from Beukers et al (2023). The task varies along three orthogonal dimensions.…”