CAREER: Extracting principles of neural computation from large scale neural recordings through neural network theory and high dimensional statistics

职业：通过神经网络理论和高维统计从大规模神经记录中提取神经计算原理

• 批准号: 1845166
• 负责人: Surya Ganguli
• 金额: $ 50万
• 依托单位: Stanford University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2019
• 资助国家: 美国
• 起止时间: 2019-10-01 至 2024-09-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1845166&HistoricalAwards=false
• 关键词: CAREER; Extracting; principles; neural; computation

Recent technological advances now enable recordings of thousands of neurons during complex behaviors. Such experimental capabilities could potentially reveal how the brain encodes sensations, forms memories, learns tasks, makes decisions, and generates motor actions. However, there exist major obstacles to attaining a scientific understanding of how the psychological capabilities of the mind emerge from the biological wetware of the brain. First, data analytic methods are not adequate to make sense of the massive datasets currently being gathered from the brain. Second, theoretical methods are not adequate for both optimally designing large-scale neural recordings, and bridging scales from the collective biophysics of many neurons to psychological processes underlying sensations, thoughts and actions. This project will develop novel data analytic and theoretical methods to extract a conceptual understanding of how the brain gives rise to cognition. These methods will be tested in large-scale recordings from many experimental labs studying perception, memory, learning, decision making and motor control. They will also be applied to developing better learning protocols and neural prosthetic devices.This project will pursue three overarching aims. It will build on advances in high dimensional statistics to develop a theory of when and how subsets of neurons reflect the collective dynamics of the much larger unobserved circuit in which they are embedded. This theory will provide quantitative guidance for the efficient design of future large-scale recording experiments. Second, it will build on advances in deep learning to develop algorithmic methods for extracting a conceptual understanding of how complex neural networks solve tasks. These algorithmic methods will elucidate which aspects of network connectivity and dynamics are essential to understanding how neural circuits perform their computations, thereby providing guidance for what to measure in future neuroscience experiments. Finally, it will advance theories of neural network learning to better understand how the structure of prior experience determines learned neural connectivity, and how this learning process can be optimized. These general theoretical advances will be refined and tested in specific, close experimental collaborations, involving: identifying feedback control laws in motor cortex, finding signatures of attractor dynamics in the hippocampal memory circuits, understanding the neural algorithms for perception in the retina and decision making in prefrontal cortex, and developing frameworks for understanding rapid rodent learning built upon prior experiences.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.

最近的技术进步使得在复杂的行为中记录数千个神经元成为可能。这种实验能力可能潜在地揭示大脑是如何编码感觉、形成记忆、学习任务、做出决定和产生运动动作的。然而，要科学地理解心灵的心理能力是如何从大脑的生物湿件中产生出来的，还存在着重大障碍。首先，数据分析方法不足以理解目前从大脑收集的大量数据集。其次，理论方法不足以优化设计大规模的神经记录，也不足以将许多神经元的集体生物物理学与潜在的感觉、思想和行动的心理过程联系起来。该项目将开发新的数据分析和理论方法，以提取对大脑如何产生认知的概念性理解。这些方法将在许多研究感知、记忆、学习、决策和运动控制的实验实验室的大规模录音中进行测试。它们还将应用于开发更好的学习协议和神经假体设备。该项目将追求三个总体目标。它将以高维统计学的进步为基础，发展一种理论，研究神经元子集何时以及如何反映它们所嵌入的更大的未被观察到的电路的集体动态。该理论将为今后大规模记录实验的高效设计提供定量指导。其次，它将以深度学习的进步为基础，开发算法方法，以提取对复杂神经网络如何解决任务的概念性理解。这些算法方法将阐明网络连接和动态的哪些方面对于理解神经回路如何执行其计算是必不可少的，从而为未来神经科学实验的测量提供指导。最后，它将推进神经网络学习理论，以更好地理解先验经验的结构如何决定学习神经连通性，以及如何优化这一学习过程。这些普遍的理论进步将在具体的、密切的实验合作中得到完善和测试，包括：识别运动皮层的反馈控制规律，在海马记忆回路中寻找吸引子动力学的特征，理解视网膜感知和前额叶皮层决策的神经算法，以及建立在先前经验基础上的理解啮齿动物快速学习的框架。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

项目成果

A unified theory for the origin of grid cells through the lens of pattern formation

• 发表时间: 2019
• 作者: Ben Sorscher;Gabriel C. Mel;S. Ganguli;Samuel A. Ocko

Universality and individuality in neural dynamics across large populations of recurrent networks.

大量循环网络中神经动力学的普遍性和个体性。

• 发表时间: 2019
• 期刊: Advances in neural information processing systems
• 作者: Maheswaranathan,Niru;Williams,AlexH;Golub,MatthewD;Ganguli,Surya;Sussillo,David
• 通讯作者: Sussillo,David

Deep learning versus kernel learning: an empirical study of loss landscape geometry and the time evolution of the Neural Tangent Kernel

• 发表时间: 2020-10
• 期刊: ArXiv
• 作者: Stanislav Fort;G. Dziugaite;Mansheej Paul;Sepideh Kharaghani;Daniel M. Roy;S. Ganguli

Reverse engineering recurrent networks for sentiment classification reveals line attractor dynamics

• 发表时间: 2019-06
• 期刊: Advances in neural information processing systems
• 作者: Niru Maheswaranathan;Alex H. Williams;Matthew D. Golub;S. Ganguli;David Sussillo

Enhancing Associative Memory Recall and Storage Capacity Using Confocal Cavity QED

• DOI: 10.1103/physrevx.11.021048
• 发表时间: 2020-09
• 期刊: arXiv: Quantum Physics
• 作者: Brendan P. Marsh;Yudan Guo;Ronen M. Kroeze;S. Gopalakrishnan;S. Ganguli;Jonathan Keeling;B. Lev