Adapting Hierarchical Circuits from Planning to Language with Computational Modeling

这一提议的目标是使用计算模型来检验存在于 PFC-BG区域内共享的神经机制，它是两个抽象规划的基础 物理域(即运动动作选择)和抽象语言处理。了解神经 人类语言的基础是认知神经科学的圣杯之一。最近，强大的模特们 从人工智能，特别是自然语言处理(NLP)，已经显示出令人印象深刻的 能够捕捉语言结构的许多方面，但还不清楚这些是如何(如果有的话)的 模型可以与人脑的处理过程相媲美。该项目将使用来自 最近在NLP(特别是变压器架构和迁移学习)方面的工作，以开发一种新的 PFC-BG区域的计算模型。在试点之后，这项研究的长期目标是 计划是使用这些计算模型产生的洞察力来改进诊断和 治疗抽象计划和语言能力受损的精神疾病--例如， 强迫症(计划障碍影响语篇连贯)和阿尔茨海默氏症(障碍 影响语法和语义之间的交互)。 这项工作被组织成三个具体目标。AIM 1将评估基于变压器的模型是否可以 解释人类在运动规划领域的认知工作记忆任务的行为数据。目标 2将评估相同的模型架构是否能够支持语言结构的学习 抽象语法。目标3将评估计算机制是否可以从 从目标1的运动规划任务到目标2的语言任务。这里，“迁移”将由 通过在电路上构建语言处理模型而提高了样本效率 它以前是专门用于运动规划任务的，而不是从头开始训练。 综上所述，拟议中的实验结果将揭示出基本的 认知能力(即运动规划)和抽象语言依赖于共享的PFC-BG电路，或者 相反，这些任务依赖于大脑中结构相似但物理上不同的网络

The goal of this proposal is to use computational modeling in order to test the hypothesis that there exists a shared neural mechanism within the PFC-BG region which underlies both abstract planning in the physical domain (i.e., motor action selection) and abstract language processing. Understanding the neural basis for human language is one of the holy grails of cognitive neuroscience. Recently, powerful models from artificial intelligence, specifically natural language processing (NLP), have demonstrated impressive ability to capture many aspects of linguistic structure, but it is not yet understood how (if at all) these models are comparable to processing in the human brain. This project will use modeling insights from recent work in NLP (specifically, Transformer architectures and transfer learning) in order to develop a new computational model of the PFC-BG region. Following the pilot, the long term goal of this research program is to use the insights generated by these computational models to improve diagnosis and treatment of psychiatric illness in which both abstract planning and language abilities are impaired–e.g., OCD (in which impairments to planning affect discourse coherence) and Alzheimer’s (in which impairments affect interaction between syntax and semantics). The work is organized into three specific aims. Aim 1 will evaluate whether Transformer-based models can account for human behavioral data on cognitive working memory tasks in the motor planning domain. Aim 2 will evaluate whether the same model architectures can support learning of linguistic structure using abstract grammars. Aim 3 will evaluate whether computational mechanisms can be transfered from the motor planning task from Aim 1 to the language task from Aim 2. Here, “transfer” will be quantified by the increase in sample efficiency that results from building the language processing model on top of circuitry that has previously been specialized for the motor planning task, as opposed to training from scratch. Taken together, the results of the proposed experiments will reveal whether it is plausible that basic cognitive abilities (i.e., motor planning) and abstract language depend on shared PFC-BG circuitry, or rather that the tasks depend on structurally similar but physically distinct networks within the brain