Interpretable machine learning for understanding the neural control of movement

现代机器学习可以拟合复杂的模型，做出准确的预测。然而，用标准的机器学习来理解大脑是一个巨大的挑战。一个关键的障碍是，通常很难理解这些模型的内部工作原理。因此，神经科学非常需要更可解释的机器学习方法，其中模型的内部工作与神经结构或功能有关。拟议的研究将开发可解释的机器学习方法，用于研究运动的神经控制。通过与Churchland博士实验室的密切合作，这些方法将被应用于回答运动控制中的三个核心问题。在前两个目标中，该研究调查了如何灵活地控制脊髓运动神经元，以及运动皮层的活动如何在广泛的运动中控制肌肉活动。除了增加科学知识外，回答这两个问题还可以增强脑机接口，以恢复对运动障碍患者的控制。在第三个目标中，该研究调查了多个神经群体如何相互作用以规划和产生运动。这个问题与许多神经系统疾病有关，其中大脑区域之间的通信中断。在该奖项的K99阶段，发生在哥伦比亚充满活力的神经科学界，申请人将接受额外的培训，在计算技术博士。这项培训将为申请人成功完成这项工作提供必要的经验，并成功地在机器学习和运动神经控制的界面上建立一个独立的实验室。

Modern machine learning can fit complex models that make accurate predictions. However, understanding the brain with standard machine learning is a great challenge. A critical impediment is that it is often difficult to understand the inner workings of these models. There is thus a great need for more interpretable machine learning methods for neuroscience, in which the inner workings of the model relate to neural structure or function. The proposed research will develop interpretable machine learning methods for studying the neural control of movement. Through close collaboration with the experimental lab of Dr. Churchland, these methods will be applied to answer three central questions in motor control. In the first two aims, the research investigates how flexibly spinal motor neurons can be controlled, and how activity in motor cortex controls muscle activity across a wide range of movements. Beyond increasing scientific knowledge, answering these two questions could enhance brain computer interfaces for restoring control to those with motor impairments. In the third aim, the research investigates how multiple neural populations interact with each other to plan and generate movement. This question is relevant to many neurological disorders, in which communication between brain areas is disrupted. During the K99 phase of this award, occurring within Columbia’s vibrant neuroscience community, the applicant will receive additional training in computational techniques from Drs. Paninski and Cunningham, and additional experimental training from Dr. Churchland. This training will provide the necessary experience for the applicant to be successful in this work and to successfully start an independent lab at the interface of machine learning and the neural control of movement.