Elucidating Principles of Sensorimotor Control using Deep Learning

项目摘要 分布式神经回路如何从复杂的肌肉骨骼系统驱动有目的的运动？这 对原则性神经刺激的理解和描述对于将其应用于 研究神经回路干扰对行为的影响，反之亦然 在行为扰动期间对神经活动的预测。这个大脑的研究目标是 倡议提案是开发受生物启发、目标和数据驱动的人工智能方法，以 阐明感觉运动控制的神经动力学基础。这项研究计划的成果将 从根本上影响我们对感觉运动控制背后的神经回路的理解。这些工具 本文中的研究成果将分发给科学界免费使用。 通过这项大脑倡议提案，我们将阐明感觉运动控制的原理，通过将 以简明和可解释的生物启发模型记录的神经数据之间的关系 测量的生物数据和相应的行为。在目标1和目标2中，我们将设计和传播一个 综合建模框架，将大规模神经和行为数据与基于物理的 肌肉骨骼系统和神经解剖学约束的建模。神经数据和神经解剖学 约束将被合并到递归神经网络模型中，该模型将通过以下方式实现预期行为 使用尖端机器学习方法的基于生物物理学的肌肉骨骼模型。建模 这里创建的框架将为在Aim 3中设计虚拟实验室提供急需的机会，以测试 神经刺激在反馈环境中的效果，并预测看不见的行为条件和 对由此产生的神经活动的行为扰动。 硬件和机器学习技术的飞速发展使我们的能力得到了极大的提高 对大规模多区域神经数据进行记录和建模。我们广泛的研究目标是推动当前的 通过结合大规模测量和测量对运动的神经控制进行建模的最新技术 将生物约束转化为感觉运动控制的理论模型。这是向(A)澄清迈出的关键一步 不同大脑区域的神经活动在控制复杂行为中的计算作用，(B)允许 美国进一步完善基于数据的运动生成理论模型，以及(C)了解哪里和 如何刺激大脑，以便有效地应用神经刺激来实现预期的行为。与 随着这些工具的开发和传播，我们希望进入一个虚拟实验室不仅仅是 分析以前进行的实验的方法，但被集成到实验管道中，以便它们 可以在大规模神经科学实验中充分利用它们的潜力。

Project Summary How do distributed neural circuits drive purposeful movements from the complex musculoskeletal system? This understanding and characterization will be critical towards the application of principled neurostimulation to specific brain regions to study the effect of neural circuit perturbations on behavior, and conversely towards predictions of the neural activity during perturbations in the behavior. The research objective of this BRAIN Initiative proposal is to develop biologically-inspired goal- and data- driven artificial intelligence methods to elucidate the neurodynamical basis of sensorimotor control. The outcomes of this research program will fundamentally impact our understanding of the neural circuits underlying sensorimotor control. The tools developed herein will be disseminated for free use by the scientific community. Through this BRAIN Initiative proposal, we will elucidate principles of sensorimotor control by incorporating recorded neural data in succinct and interpretable biologically-inspired models of the relationships between the measured biological data and the corresponding behavior. In Aims 1 and 2, we will design and disseminate a comprehensive modeling framework that integrates large-scale neural and behavioral data with physics-based modeling of the musculoskeletal system and neuroanatomical constraints. The neural data and neuroanatomical constraints will be incorporated in recurrent neural network models that will achieve desired behavior through biophysically-based musculoskeletal models using cutting-edge machine learning methods. The modeling framework created here will provide a much-needed opportunity to design a virtual laboratory in Aim 3 that tests the effect of neural stimulation in a feedback setting, and predicts the effect of unseen behavioral conditions and behavioral perturbation on resulting neural activity. Breakneck advances in hardware and machine learning techniques have led to vast improvements in our ability to record and model large-scale multi-regional neural data. Our broad research goal is to advance the current state-of-the-art for modeling the neural control of movements by incorporating large-scale measurements and biological constraints into theoretical models of sensorimotor control. This is a critical step towards (a) elucidating the computational role of neural activity from different brain regions in controlling complex behavior, (b) allowing us to further refine theoretical models of movement generation based on data, and (c) understanding where and how to stimulate the brain in order to efficiently apply neurostimulation for achieving desired behavior. With the development and dissemination of these tools, we hope to enter an era where virtual laboratories are not just a way to analyze previously performed experiments, but are integrated into experimental pipelines such that they can be utilized to their full potential during large-scale neuroscience experiments.