CAREER: Elucidating principles of cortical computation with recurrent neural networks

职业：利用循环神经网络阐明皮质计算原理

• 批准号: 1943467
• 负责人: Jonathan Kao
• 金额: $ 57.5万
• 依托单位: University of California-Los Angeles
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-10-01 至 2025-09-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1943467&HistoricalAwards=false
• 关键词: CAREER; Elucidating; principles; cortical; computation

The goal of this project is to develop a machine learning based modeling framework to discover new mechanisms of brain function, with a focus on how the brain supports decision-making and movement. Understanding the brain's inner workings, including how circuits of neurons compute and give rise to these behaviors, is critical for better diagnosing and treating cognitive and motor disorders. A challenge to studying brain function is its complexity: billions of neurons across multiple brain areas interact in coordinated ways. This project gains new insight into brain function by modeling these coordinated and multi-area computations with deep neural networks. These networks, unlike the brain, are fully observed: all artificial neurons, their activity, and their connections are known. As such, neural networks that are trained to compute like the brain can be analyzed to discover new mechanistic insights for brain function. This project will also use these insights to develop higher performance brain-computer interfaces that help the paralyzed by decoding thoughts into actions.This project will use recurrent neural networks as in silico models of brain areas. New neural network architectures will be trained to do the same tasks and behaviors that animals perform in experimental labs. Critically, these neural networks will be trained to harness information from basic neuroscience including anatomy and neuron recordings, so that its artificial neurons resemble real neurons. After training, neural networks will be analyzed to propose new computational mechanisms for how populations of neurons compute to produce behaviors. New hypotheses of brain function from these networks will be tested in collaboration with experimental labs. These insights and models will be incorporated into new algorithms for brain-computer interfaces that aim to better decode one's intentions from his or her neural activity. The outcomes of this project will also be translated into educational and outreach materials to reach a broad audience, contributing to the training of a next generation of computational neuroscientists and neural engineers.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

该项目的目标是开发一个基于机器学习的建模框架，以发现大脑功能的新机制，重点是大脑如何支持决策和运动。了解大脑的内部运作，包括神经元回路如何计算和产生这些行为，对于更好地诊断和治疗认知和运动障碍至关重要。研究大脑功能的一个挑战是它的复杂性：跨多个大脑区域的数十亿神经元以协调的方式相互作用。该项目通过使用深度神经网络对这些协调和多区域计算进行建模，从而获得对大脑功能的新见解。与大脑不同，这些网络是完全可以观察到的：所有人工神经元，它们的活动和它们的连接都是已知的。因此，可以分析经过训练以像大脑一样计算的神经网络，以发现大脑功能的新机制见解。该项目还将利用这些见解开发更高性能的脑机接口，通过将思想解码为行动来帮助瘫痪者。该项目将使用循环神经网络作为大脑区域的计算机模型。新的神经网络架构将被训练来完成动物在实验室中执行的相同任务和行为。至关重要的是，这些神经网络将被训练来利用基础神经科学的信息，包括解剖学和神经元记录，使其人工神经元类似于真实的神经元。 训练后，将分析神经网络，为神经元群体如何计算产生行为提出新的计算机制。来自这些网络的大脑功能的新假设将与实验室合作进行测试。这些见解和模型将被纳入脑机接口的新算法中，旨在更好地从他或她的神经活动中解码一个人的意图。该项目的成果也将被转化为教育和推广材料，以达到广泛的受众，有助于下一代计算神经科学家和神经工程师的培训。该奖项反映了NSF的法定使命，并已被认为值得通过使用基金会的智力价值和更广泛的影响审查标准进行评估的支持。

项目成果

A mechanistic multi-area recurrent network model of decision-making

决策的机械多区域循环网络模型

• 发表时间: 2021
• 期刊: Advances in neural information processing systems
• 作者: Kleinman M;Chandrasekaran C*;Kao JC*
• 通讯作者: Kao JC*

Usable Information and Evolution of Optimal Representations During Training

训练期间的可用信息和最佳表示的演变

• 发表时间: 2021
• 期刊: International Conference on Learning Representations (ICLR
• 作者: Kleinman, Michael;Achille, Alessandro;Idnani, Daksh;Kao, Jonathan C
• 通讯作者: Kao, Jonathan C

Learning rule influences recurrent network representations but not attractor structure in decision-making tasks

学习规则影响循环网络表示，但不影响决策任务中的吸引子结构

• 发表时间: 2021
• 期刊: Advances in neural information processing systems
• 作者: McMahan B;Kleinman M;Kao JC
• 通讯作者: Kao JC