美国哥伦比亚大学C. Daniel Salzman、Stefano Fusi等研究人员合作揭示海马和前额叶皮层的抽象几何。相关论文于2020年10月14日在线发表在《细胞》杂志上。

研究人员表征了猴子执行不同隐藏和显式变量描述任务的神经表现。抽象通过在操作上使用神经解码器的泛化性能在不用于训练的任务条件下被定义，这需要特定的神经表示几何形状。前额叶皮层、海马和模拟的神经网络中的神经集合同时表示反映抽象的几何形状中的多个变量，但仍允许线性分类器解码大量其他变量（高破碎维数）。此外，这种几何形状相对于任务事件和性能也发生了变化。

这些发现阐明了大脑和人工系统如何以抽象形式表示变量，同时保留了高破碎维度所赋予的优势。

据介绍，维度因素困扰着强化学习和决策制定的模型。抽象过程通过构造描述不同实例共享的特征的变量来解决此问题，从而减少维度并在新颖情况下实现泛化。

附：英文原文

Title: The Geometry of Abstraction in the Hippocampus and Prefrontal Cortex

Author: Silvia Bernardi, Marcus K. Benna, Mattia Rigotti, Jérme Munuera, Stefano Fusi, C. Daniel Salzman

Issue&Volume: 2020-10-14

Abstract: The curse of dimensionality plagues models of reinforcement learning and decisionmaking. The process of abstraction solves this by constructing variables describingfeatures shared by different instances, reducing dimensionality and enabling generalizationin novel situations. Here, we characterized neural representations in monkeys performinga task described by different hidden and explicit variables. Abstraction was definedoperationally using the generalization performance of neural decoders across taskconditions not used for training, which requires a particular geometry of neural representations.Neural ensembles in prefrontal cortex, hippocampus, and simulated neural networkssimultaneously represented multiple variables in a geometry reflecting abstractionbut that still allowed a linear classifier to decode a large number of other variables(high shattering dimensionality). Furthermore, this geometry changed in relation totask events and performance. These findings elucidate how the brain and artificialsystems represent variables in an abstract format while preserving the advantagesconferred by high shattering dimensionality.

DOI: 10.1016/j.cell.2020.09.031

Source: https://www.cell.com/cell/fulltext/S0092-8674(20)31228-9