The virtual rodent: a platform to study the artificial and biological control of natural behavior

项目摘要 在不确定的环境中控制复杂的物体是我们的大脑进化到完美的挑战，然而 实现灵活和稳健控制的算法和神经网络实现一直很难 辨认身份。这一建议的前提是这样一种想法，即通过拥抱复杂的 潜在的控制系统，包括它们控制的身体和动物行为的多样性。为了测试这一点 Idea，更广泛地说，提供了一个通用平台来询问神经电路级原理和 具体的运动控制机制，我提出了虚拟啮齿动物。这种硅胶动物将会有一个 身体像真正的老鼠，体验正常的物理，并接受训练以产生自然主义的老鼠行为。它将会有 一种可以完全被审问、操纵和重新配置的人工大脑。在建立了这个平台之后，我 将开发一种分析方法来比较从自由移动的动物到网络的体内神经活动 模型的表示形式。这一努力扩展了最近将神经表示联系起来的方法 与感官系统中任务优化的人工模型的表示，使得能够比较神经 在运动领域和复杂行为中使用分析模型的活动。然后，我提议进一步 开发虚拟啮齿动物，探索与动物的分级控制和运动学习有关的问题 机器。 在这项研究的F99阶段，我将继续开发虚拟啮齿动物作为研究的平台 对自然行为的人工和生物控制。具体地说，在目标1中，我将最终确定一个行为 测量、处理和建模管道，以训练人工神经网络来模拟 在物理模拟器中的真实啮齿动物，验证其性能，并演示其作为 内置马达控制。在目标2中，我将从真实啮齿动物的运动中心记录它们自由移动和 将他们的神经活动与执行相同不同动作的模型的网络活动进行比较。 在这项研究的K00阶段，我将在虚拟啮齿动物模型的基础上进行扩展，以研究分层 控制，这是灵活和适应性哺乳动物控制的保守特征。我将训练一个人工神经网络来 重用作为F99阶段的一部分创建的低级控制模块，以自主地解决通常的电机任务 用于运动神经科学研究。这一目标将对运动神经科学领域具有重大价值。 便于比较执行受控任务的动物的神经活动与 执行相同任务的物理模拟模拟的分析模型。总而言之，这些目标提供了一个 研究运动的神经控制的新路径，一条包含行为和 生物力学，以促进我们对健康和疾病中灵活和适应性运动控制的理解。

Project Summary Controlling complex bodies in uncertain environments is a challenge our brains have evolved to perfect, yet the algorithms and neural network implementations that enable flexible and robust control have been difficult to identify. This proposal is premised on the idea that progress will be served by embracing the complexities of the underlying control systems, including the bodies they control and the diversity of animal behavior. To test this idea and, more generally, provide a versatile platform for interrogating the neural circuit-level principles and mechanisms underlying embodied motor control, I propose the virtual rodent. This in-silico animal will have a body like a real rat, experience normal physics, and be trained to produce naturalistic rat behaviors. It will have an artificial brain that can be fully interrogated, manipulated, and reconfigured. After establishing this platform, I will develop an analysis approach to compare in-vivo neural activity from freely moving animals to the network representations of the model. This endeavor expands upon recent approaches linking neural representations with the representations of task-optimized artificial models in sensory systems, enabling the comparison of neural activity with analytical models in the motor domain and during complex behavior. I then propose to further develop the virtual rodent to probe questions related to hierarchical control and motor learning in animals and machines. In the F99 phase of this proposed research, I will continue to develop the virtual rodent as a platform to study the artificial and biological control of natural behavior. Specifically, in Aim 1, I will finalize a behavioral measurement, processing, and modeling pipeline to train artificial neural networks to imitate the behaviors of real rodents while in a physical simulator, validate its performance, and demonstrate its utility as a model for embodied motor control. In Aim 2, I will then record from motor centers of real rodents as they freely move and compare their neural activity to the network activity of models enacting the same diverse movements. In the K00 phase of this proposed research, I will expand upon the virtual rodent model to study hierarchical control, a conserved feature of flexible and adaptive mammalian control. I will train an artificial neural network to reuse lower-level control modules created as part of the F99 phase to autonomously solve motor tasks commonly used in motor neuroscience research. This Aim is of great value to the field of motor neuroscience as it will facilitate the comparison of neural activity of animals performing controlled tasks with the network activity of analytical models performing physically simulated analogues of the same tasks. Together, these Aims offer a new path in the study of the neural control of movement, one which embraces the complexity of behavior and biomechanics to advance our understanding of flexible and adaptive motor control in health and disease.