Ghost in the Machine: Melding Brain, Computer and Behavior

植入式设备在神经护理中发挥着更大的作用，但它们的有效性 有限，因为他们对人类的想法、感觉和行为视而不见--大多数因素 极大地影响我们的健康。将外围传感器连接到植入物可能会有所帮助，但这不是很有帮助吗 如果这些设备直接问我们会不会更容易？有了这些知识，下一代机器将会 有效地将大脑中的神经活动驱动到健康状态。他们还会很快学会行为 这会恶化健康，并引导我们做出更好的选择。尽管DARPA、NIH和Neuralink 花费数百万美元购买脑机接口的新硬件，没有人关注 主机和机器之间相互的、自然的交流。人们迫切需要 使设备能够学习和指导人类行为的新颖、实用的方法。 在这项应用中，我提出了开发新一代自主脑机 界面-可以提问、记录、行动的设备-并将应用于以下各项的学习算法结合起来 神经信号与它们的人类宿主的教导。有了这些植入物的生活将需要一个微妙的人类- 机器对话，在这种对话中，设备和人类相互教授和学习。人类会通知 关于我们正在做什么和感觉的智能算法，而机器将结合这一点 将信息转化为治疗，并引导我们以个性化的方式优化生活质量。这是一个 从今天的简单设备转变为范式，这些设备由医生在偶然情况下进行编程 办公室来访。我建议以一种实用的、可伸缩的方式演示此范例，使用当前 癫痫植入物可以迅速翻译成许多神经疾病。 为了实现这一目标，我将以新颖的方式融合几项尖端技术，包括： (1)最先进的高带宽植入物，可以采样神经活动，连接到巨大的云- 基于计算能力来处理它，并干预以调节大脑、脊髓或外周 神经活动。这项工作利用了我过去20年的经验；(2)我将部署强大的 以新颖的方式使用新的计算机科学工具。我将使用卷积神经网络(也称为深沉 学习)为了从海量的连续高带宽神经数据流中学习模式，构建 使用自然语言处理(NLP)的双向人机界面，以及Probe 网络与人类行为和电刺激的变化，并引导干预措施 使用强化学习的治疗目标。结合这些计算机科学，机器 对我来说，人类行为的学习技术和衡量标准是一个新的研究领域 这将利用我在临床神经学和工程学方面的独特背景来建立一个新的班级 交互的人类治疗设备。

Implantable devices are playing a greater role in neurologic care, but their effectiveness is limited, because they are blind to human thoughts, feelings, and behavior – factors that most dramatically affect our health. Coupling peripheral sensors to implants might help, but wouldn’t it be easier if the devices just asked us? Armed with this knowledge, next generation machines will more effectively drive neural activity in the brain to healthy states. They will also quickly learn behaviors that worsen health and guide us to better choices. Though DARPA, the NIH, and Neuralink are spending millions of dollars on new hardware for brain-computer interfaces, none focus on reciprocal, natural communication between host and machine. There is a desperate need for novel, practical methods that enable devices to learn from and guide human behavior. In this application I propose to develop a new generation of autonomous brain-machine interfaces – devices that can question, record, act - and combine learning algorithms applied to neurosignals with teaching by their human hosts. Life with these implants will entail a subtle human- machine dialogue in which devices and humans teach and learn from each other. Humans will inform intelligent algorithms about what we are doing and feeling, while machines will incorporate this information into therapy and guide us to optimize quality of life in personalized ways. This is a paradigm shift from today’s simple devices, which are programmed by physicians during occasional office visits. I propose to demonstrate this paradigm in a practical, scalable way using current epilepsy implants that is rapidly translatable to many neurological disorders. To achieve this goal, I will meld several cutting-edge technologies in novel ways, including: (1) State-of-the-art, high bandwidth implantables that sample neural activity, link to vast cloud- based computational power to process it, and intervene to modulate brain, spinal cord or peripheral neural activity. This work utilizes my experience from the past 20 years; (2) I will deploy powerful new computer science tools in novel ways. I will use convolutional neural nets (a.k.a. Deep Learning) to learn patterns from vast streams of continuous high-bandwidth neural data, build a two way human-machine interface using Natural Language Processing (NLP)., and probe networks with changes in human behavior and electrical stimulation and guide interventions toward therapeutic goals using Reinforcement Learning. Combining these computer science, machine learning techniques and measurements of human behavior is a new area of investigation for me that will leverage my unique background in clinical neurology and engineering to build a new class of interactive, human therapeutic devices.