The Emergence of Abstract Representations in Learning and Development

抽象的知识，如类别，使人类的大脑能够以一种有效的方式处理来自环境的大量信息，即通过降低信息的复杂性和维度。人类认知的一个方面是人工智能（AI）无法比拟的，那就是在经历少量样本后进行泛化的能力。例如，在经历了几次鸟的例子后，孩子可能会产生“鸟”的概念。抽象知识的获取被认为需要抽取和巩固事件记忆（即情景记忆）的规律。另一方面，情景记忆通过将一个情节的独特特征组合在一起来保留单个事件的特殊性。到目前为止，人们还没有很好地理解大脑是如何学会捕捉不同经验的规律以获得分类知识的，以及这些动态过程在儿童和成人之间有何不同。因此，本项目将研究类别抽象表征的出现，以及它如何与特定经历的情景记忆相互作用。我们将为儿童和成人开发一个适合年龄的实验范式，以跟踪人类大脑中抽象知识和情景记忆的表征。这些表征不仅可以通过行为测量来索引，还可以通过高时间和空间分辨率的测量来索引，例如眼动追踪和功能磁共振成像数据，以表征跨越时间的动态过程。计算方法将用于定量表征多变量、高维数据中的底层结构，同时，探索人工智能模型在多大程度上可以在潜在的表征结构中近似人类数据。了解分类知识是如何在人类大脑中出现的，特别是在大脑可塑性增强的儿童发育过程中，可能有助于为未来的人工智能模型提供信息，这些模型能够在很少的学习事件中有效地学习新类别，并随后将其灵活地应用于新情况。

Abstract knowledge, such as categories, enables the human brain to process the overwhelming amount of incoming information from the environment in an efficient way, namely by reducing the complexity and dimension of information. One aspect of human cognition unmatched by artificial intelligence (AI) is the capacity to generalize after experiencing few samples. For example, a concept for “bird” as a category may emerge in a child after experiencing just several instances of birds. Acquisition of abstract knowledge is assumed to entail the extraction and consolidation of regularities across event memories (i.e. episodic memory). On the other hand, episodic memory preserves the specificity of individual events by binding together unique combinations of features from an episode. Thus far, it is not well understood how the brain learns to capture regularities across experiences for the acquisition of categorical knowledge, and how these dynamic processes differ between children and adults. Therefore, this project will examine the emergence of abstract representation for categories and how it interacts with episodic memory of specific experiences. We will develop an age appropriate experimental paradigm for children and adults that allows tracking the representations of both abstract knowledge and episodic memory in the human brain. These representations are indexed not just by behavioral measures, but also measures with high temporal and spatial resolution, i.e. eye-tracking and functional magnetic resonance imaging data, in order to characterize the dynamic processes across time. Computational methods will be used to quantitatively characterize the underlying structures within the multivariate, high-dimensional data, and at the same time, explore the extent to which AI models can approximate the human data in the latent representational structure. Gaining insights on how categorical knowledge emerges in the human brain, particularly during child development with its heightened brain plasticity, may help to inform future AI models that are capable of learning new categories efficiently with few learning episodes and subsequently applying them flexibly to new situations.