Representation Learning for Continual Task Performance

人类认知的一个标志是连续学习和执行多个任务的能力（“持续”学习）。心理学家和神经科学家研究了执行功能如何控制任务之间的转换。然而，在这些研究中(1)通常为受试者提供完整的先验指令(2)刺激高度简化。相比之下，在现实世界中，(1)视觉信息是高维的，(2)需要在没有指令的情况下通过试错来检测最佳刺激-反应偶然性，(3)大脑面临的挑战是，以一种避免干扰未来学习的方式来表示任务集，并能够在类似的环境中灵活地重用神经代码。然而，这种持续学习的神经基础仍然是个谜。值得注意的是，在机器学习中，即使是最先进的算法也无法持续获取世界的结构化、受保护和可重用的表示。本文的目的是通过以下三个实验来研究持续学习的计算机制和神经表征：实验1：持续任务学习的神经关联。我研究了与交错学习相比，顺序学习是否促进了人类PFC中任务集的多个压缩表示的出现，从而防止了干扰。实验二：持续任务绩效的规则迁移。大脑如何在部分与先前经验相关的新情况下重用抽象知识？实验3：分类的无监督结构学习。对环境统计结构的低级感觉学习是否为以后的任务学习提供了支撑？

A hallmark of human cognition is the ability to learn and execute multiple tasks in succession ("continual" learning). Psychologists and Neuroscientists have investigated how executive functions control switching between tasks. However, in these studies (1) subjects were usually provided with full instructions a priori and (2) stimuli were highly simplified. In contrast, in the real world, (1) visual information is high-dimensional, (2) optimal stimulus-response contingencies needto be detected without instructions via trial and error and (3) the brain is challenged to represent task-sets in a way that avoids interference with future learning and enables flexible reuse of neural codes in similar environments. The neural basis of this continual learning, however, remains enigmatic. Notably, in Machine Learning, even state-of-the art algorithms fail to continuously acquire structured, protected and reusable representations of the world. The aim of the proposedDPhil is to investigate the computational mechanisms and neural representations underlying continual learning via the following three experiments:Experiment 1: Neural Correlates of Continual Task Learning.I investigate if sequential in contrast to interleaved learning facilitates the emergence of multiple compressed representations of task-sets in human PFC that are protected against interference.Experiment 2: Rule Transfer for Continual Task Performance.How does the brain reuse abstract knowledge in novel situations that are partially related to previous experience? Experiment 3: Unsupervised Structure Learning for Categorisation.Does low-level sensory learning of the statistical structure of our environment serve as scaffold for later task learning?