Navigating Knowledge Spaces Using Cognitive Maps

大脑最基本的能力之一是灵活地形成空间环境的认知地图以供导航。最近的研究表明，大脑中认知地图的概念远远超出了物理空间的表征，表明认知地图可以灵活地以特定任务的方式形成，并且可以包含不同抽象层次的语义表征。这种与任务相关的语义知识的有序映射可以发挥许多重要作用，包括分类、记忆和对以前没有遇到过的对象属性的外推。然而，目前还不清楚更抽象的语义属性的认知地图是如何形成的，它们如何支持类别学习或视觉搜索等任务，以及它们在不同任务之间的泛化程度如何。因此，通过利用ARENA研究小组的协作力量和科学范围，本文提出的研究旨在通过开发新的人工智能模型并在人脑中测试其预测来揭示认知地图的形成和计算效益。首先，我们将使用最先进的深度生成对抗网络（GANs）为抽象表示建模认知地图。gan可以产生新颖的、接近真实的数据，如复杂的自然图像，其潜在空间包含有序的、语义上有意义的表示。基于这些属性，我们提出了从gan的潜在空间构建包含视觉对象特征语义特征的低维可解释映射的方法。然后，我们将此方法应用于有限输入数据的类别学习问题。由于gan可以模拟接近真实的数据，因此它们可以用于数据增强，并且我们探索了在适当的认知地图指导下，gan对数据进行采样的能力如何使基于少量可用数据的类别学习更有效。采样数据的过程类似于心理意象，因此我们假设由认知地图引导的采样可以作为人类类别学习的机制。该模型预测了类别学习期间大脑活动高维结构的变化，我们在ARENA进行的平行实验中测试了这些预测。最后，通过在ARENA联盟内的紧密合作，我们将分析认知地图的概括性及其在不同任务中的灵活性，我们将在类别学习和使用与人类行为实验密切相关的计算模型学习场景语法的背景下对此进行研究。通过研究认知地图的核心原理和功能意义，我们旨在更好地理解大脑和人工智能系统如何以特定任务的方式高效地学习抽象知识。

One of the most fundamental abilities of the brain is to flexibly form a cognitive map of the spatial environment for navigation. Recent work suggests that the concept of cognitive maps in the brain extends far beyond the representation of physical space by demonstrating that cognitive maps can form flexibly and in a task-specific manner and can encompass semantic representations at various levels of abstraction. Such orderly maps of task-relevant, semantic knowledge could serve a number of important roles, including categorization, memorization, and extrapolation towards object properties that were not encountered before. However, it is currently not well understood how cognitive maps of more abstract semantic properties form, how they can support tasks like category learning or visual search and to what extent they generalize across different tasks. Hence, by leveraging the collaborative strength and scientific scope of the ARENA research group, the research proposed here aims to shed new light on the formation and computational benefits of cognitive maps by developing new AI models and testing their predictions in the human brain. First, we will model cognitive maps for abstract representations using state-of-the-art deep generative adversarial networks (GANs). GANs can produce novel, near realistic data, such as complex natural images, and its latent space contains well-ordered, semantically meaningful representations. Building upon these properties, we propose methods to construct low-dimensional interpretable maps from the latent spaces of GANs that encompass characteristic semantic features of visual objects. Then, we apply this approach to the problem of category learning with limited input data. Since GANs can mimic near realistic data, they can be used for data augmentation and we explore how the GANs’ ability to sample data, when guided by suitable cognitive maps, can make category learning based on little available data more efficient. The process of sampling data is akin to mental imagery, and so we hypothesize that sampling guided by a cognitive map could serve as a mechanism for category learning in humans. The model predicts changes in the high dimensional structure of brain activity during category learning and we test these predictions in parallel experiments conducted within ARENA. Finally, through intense collaborations within the ARENA consortium we will analyze the generalizability of cognitive maps as well as their flexibility across different tasks and we will investigate this in the context of category learning and learning scene grammar using computational models that are tightly linked to behavioral experiments in humans.By studying the core principles and functional significance of cognitive maps, we aim to gain an improved understanding of how brains and AI systems can learn abstract knowledge efficiently, and in a task-specific manner.