CAREER: Bottom-Up Localized Online Learning with Spintronic Neuromorphic Networks

该奖项全部或部分由《2021年美国救援计划法案》（公法117-2）资助。人工智能（AI）和神经网络利用了人类大脑的灵感，使机器学习系统对社会产生了深远的影响。人工智能系统在系统部署后持续学习的能力尤其有前途，因为这种在线学习提供了开发新功能和适应不断变化的环境的潜力。然而，传统的机器学习算法需要对大型数据集进行大量的数学运算，需要复杂的硬件和大量的能源消耗，这阻碍了部署后在线学习的人工智能系统的发展。因此，该项目建议从神经生物学中获得进一步的灵感，利用从局部突触活动中产生的节能在线学习算法。这种本地化学习方法将极大地推动在线学习系统的发展，影响自动驾驶汽车和健康监测设备等广泛的自主应用。该项目还将通过K-12教育推广、本科研究、研究生教育以及本地和国际社会的参与，扩大对计算机的参与。为了实现高效的在线学习，该项目将采用自下而上的方法来设计神经形态网络。与传统的自上而下的方法不同，在这种方法中，监督学习算法（如反向传播）是在计算昂贵的电路中实现的，这种自下而上的方法将人工神经元和突触相互连接，这样，从局部突触更新规则中产生节能的无监督学习算法。该项目将专注于具有模拟和滞后行为的自旋电子神经形态元件，利用最近在铸造制造能力方面取得的显著进展。特别是，从这种自下而上的方法中产生的学习算法将被数学表征，允许自旋电子神经形态学习网络的设备-电路-算法协同设计。这些自旋电子神经形态网络将通过实验证明从局部学习规则中生成有效的学习算法，并且将开发设备和系统优化目标，为转化为实际的人工智能系统提供路线图。总之，该项目将加深对自旋电子物理的认识，增加对神经系统实现学习的机制的科学理解，并为具有在线学习的革命性人工智能系统开辟一条途径。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This award is funded in whole or in part under the American Rescue Plan Act of 2021 (Public Law 117-2).Artificial intelligence (AI) and neural networks have leveraged inspiration from the human brain to enable machine-learning systems that deeply impact society. The capability of an AI system to continually learn after system deployment is particularly promising, as this online learning provides the potential to develop new functionalities and adapt to changing environments. However, conventional machine-learning algorithms require the application of an enormous quantity of mathematical operations to large data sets, requiring complex hardware and large energy consumption that hinders the development of AI systems with post-deployment online learning. This project therefore proposes taking further inspiration from neurobiology, with energy-efficient online learning algorithms that emerge from local synapse activity. This localized learning approach will significantly advance the development of online learning systems, impacting a wide range of autonomy applications such as self-driving cars and health-monitoring devices. This project will also broaden participation in computing through K-12 educational outreach, undergraduate research, graduate education, and the involvement of the local and international communities.To enable energy-efficient online learning, this project will apply a bottom-up approach to the design of neuromorphic networks. Rather than the conventional top-down approach in which supervised learning algorithms (such as backpropagation) are implemented in computationally-expensive circuits, this bottom-up approach will interconnect artificial neurons and synapses such that energy-efficient unsupervised learning algorithms emerge from localized synaptic updating rules. This project will focus on spintronic neuromorphic components with analog and hysteretic behaviors, leveraging the remarkable recent progress in foundry fabrication capabilities. In particular, the learning algorithms that emerge from this bottom-up approach will be mathematically characterized, permitting device-circuit-algorithm co-design of spintronic neuromorphic learning networks. These spintronic neuromorphic networks will be experimentally demonstrated to generate effective learning algorithms from localized learning rules, and targets for device and system optimization will be developed to provide a roadmap for translation to practical AI systems. Altogether, this project will deepen knowledge of spintronic physics, increase scientific understanding of the mechanisms through which learning is achieved by neural systems, and open a pathway for revolutionary AI systems with online learning.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.