EFRI BRAID: Neurally Inspired, Resilient Closed Loop Feedback Control of Learned Motor Dynamics

该项目解决了脑机接口(BMI)这一新兴领域中一个尚未满足的关键挑战。通过将大脑活动直接映射到运动和语音，BMI技术在恢复因受伤或疾病而丧失的运动和沟通能力方面具有巨大的潜力。然而，现有的BMI仅在特定任务和情况下运行良好，恢复能力仍远未达到手部动作和言语的正常技能水平。例如，用于恢复手运动的BMI可以在受控的实验室环境中拾取和移动物体，但当引入新物体时可能会灾难性地失败。同样，部分恢复语音和交流的BMI仅限于几十个单词的词汇，或者需要逐个字母地拼写单词。为了克服这些挑战并超越目前的最先进水平，该项目开发了BMI，它结合了最近在理解大脑如何控制复杂运动方面的重大进展。主要目标是提供更好的BMI算法，根据大脑活动预测预期的运动和语音。该项目使用基于神经科学的真实大脑活动模拟来快速制作BMI算法的原型，然后通过实验测试这些算法，以促进对大脑如何控制运动的了解。这种方法将加速BMI的发展，使运动和语言能力恢复到健全人的水平。为了扩大对计算神经科学社区的参与，主要研究人员将在研究的同时，为本科生举办研讨会，学习和应用计算神经科学技能。为了提供教育资源，并将学生与更广泛的科学界联系起来，将公开展示学生的工作以及研讨会材料。该项目的主要目标是应用了解生物运动控制的最新理论进展来开发快速有效的持续学习的BMI解码策略，从而实现跨多个行为背景的稳健控制。这些努力建立在前馈神经网络解码器的基础上，这些解码器已经在受控环境中展示了初步的成功。现有的算法将用生物马达系统的理论和经验激励的特征来增强。通过结合经过验证的模块化递归神经网络(MRNN)，在Silico中加速了BMI算法的设计。通过模拟BMI测量神经活动的运动区的大脑动力学，mRNN将用户模拟为BMI控制器。第二个目标是针对现有的非人灵长类(NHP)和鸣禽神经生理学数据评估算法创新。这些离线后的分析测试了算法在开环中推断预期电机行为的能力。第三个目标是通过在NHP和鸣禽身上的活体实验来评估和验证所开发的策略。新的实验将为更广泛的行为背景操作收集数据，以进一步测试所设计算法的学习和泛化能力。关键是，NHP实验将在闭环系统BMI控制过程中验证算法。通过共同开发跨物种和上下文的神经科学理论和算法，拟议的工作有可能开发和验证优秀的BMI解码算法，并揭示开发与具有极高能效和灵活性的智能控制器交互的机器学习算法的普遍原则。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

This project addresses a critical unmet challenge in the emerging field of brain machine interfaces (BMIs). By directly mapping brain activity to movement and speech, BMI technology holds great potential for restoring movement and communication abilities lost to injury or disease. Existing BMIs only operate well for specific tasks and situations, however, and restoration capability remains far from normal levels of skill in hand movements and speech. For example, BMIs for restoring hand movement can pick up and move objects in controlled laboratory settings, but may fail catastrophically when a new object is introduced. Likewise, BMIs that partially restore speech and communication are limited to vocabularies of only dozens of words, or require words to be spelled out letter-by-letter. To overcome these challenges and advance beyond the current state-of-the-art, this project develops BMIs that combine recent, major advances in the understanding of how the brain controls complex movement. The main goal is to deliver better BMI algorithms to predict intended movement and speech from brain activity. The project uses neuroscience-based simulations of real brain activity to rapidly prototype BMI algorithms that are then tested experimentally to advance knowledge of how the brain controls movement. This approach will accelerate the development of BMIs that restore movement and speech ability to the level of an able-bodied person. To broaden participation in the computational neuroscience community, the principal investigators will, in parallel to the research, host workshops for undergraduate students to learn and apply computational neuroscience skills. To provide educational resources and connect students to the broader scientific community, a showcase of students’ work along with the workshop materials will be made publicly available. The primary objective of this project is to apply recent theoretical advances in the understanding of biological motor control to develop BMI decoding strategies for rapid and effective continual learning, enabling robust control across multiple behavioral contexts. These efforts build upon feedforward neural network decoders that have demonstrated initial success in controlled environments. Existing algorithms will be augmented with theoretically and empirically motivated features of the biological motor system. BMI algorithm design is accelerated in silico by incorporating validated modular recurrent neural networks (mRNNs). By simulating brain dynamics in the motor areas from which the BMI measures neural activity, mRNNs emulate the user as a BMI controller. The second objective evaluates algorithm innovations against existing nonhuman primate (NHP) and songbird neurophysiology data. These offline, post hoc analyses test the algorithms’ capacity to infer intended motor behavior in open loop. The third objective evaluates and validates developed strategies with in vivo experiments in NHP and songbirds. New experiments will collect data for a wider set of behavioral context manipulations to further test the learning and generalization capacity of designed algorithms. Critically, NHP experiments will validate algorithms during closed loop BMI control. By co-developing neuroscience theory and algorithms across species and contexts, the proposed work has the potential to develop and validate superior BMI decoding algorithms and to uncover generalized principles for developing machine learning algorithms that interact with intelligent controllers with extreme energy efficiency and flexibility.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.