The emergence of abstract structure knowledge across learning and sleep

点击翻译按钮获取中文摘要

In any given cognitive domain, representations of individual elements are not independent but are organized by means of structured relations. Representations of this underlying structure are powerful because they can allow generalization and inference in novel environments. In the semantic domain, structure captures associations between different semantic features or concepts (e.g., green, wings, can fly) and is known to influence the development and deterioration of semantic knowledge. We recently found that humans find it easier to learn novel categories that contain clusters of reliably co-occurring features, revealing an influence of structure on novel category formation. However, a critical unknown is whether learned representations of structure are closely tied to category-specific elements, or whether they become abstract to some extent, transformed away from the experienced features. Further, if abstract structural representations do emerge, prior work provides intriguing hints that they may require offline consolidation during awake rest or sleep. We have developed a paradigm in which carefully designed graph structures govern the pattern of feature co-occurrences within individual categories. Here we implement a "structure transfer'' extension of this paradigm in order to determine whether learning one structured category facilitates learning of a second identically structured category defined by a new set of features. This facilitation would provide evidence that structure representations are abstract to some degree. Aim 1 will use these methods to evaluate whether abstract structural representations emerge immediately during learning. Aim 2 will determine whether these representations persist, or emerge, over a delay, and whether sleep-based consolidation in particular is needed. The role of replay of recent experience during sleep will be evaluated using electroencephalography (EEG) paired with closed-loop targeted memory reactivation (TMR), a technique that enables causal influence over the consolidation of recently learned information in humans. This work will inform and constrain theories of semantic learning as well as theories of structure learning and representation more broadly.