NCS-FO: Integrated neuroengineering of brain-inspired algorithms for parsing realistic environments

在某些任务上，现代计算机的表现远远超过了人脑——例如，大规模的数值计算，或者对有组织信息的精确回忆。但对于其他重要的任务，人类和其他动物的大脑远远优于任何已建成的计算系统，无论是在它们能做的事情方面，还是在所需能量低得惊人方面。例如，人们不仅可以通过面部特征，还可以通过步态或其他他们甚至无法清晰表达的微妙动作来识别远处熟悉的人。人类和其他动物能迅速地从环境中获取信息，而且还能在新的、不可预见的情况下聪明地应用这些信息。大脑启发计算的研究致力于学习基本的新方法来思考计算机如何处理信息，这样它们就可以在这种定义薄弱的开放世界问题上表现得更好。与此同时，物理学的进步已经产生了新的光学材料和方法，这些材料和方法可以非常快速地执行计算，而且能耗非常低——如果我们感兴趣的问题能够以与这些大脑启发的计算技术兼容的方式构建起来的话。该项目旨在开发一种受大脑启发的计算网络，该网络能够快速学习并解决一系列现实世界的识别任务，并将该网络部署到便携式设备以及使用这些先进物理基板构建的定制测试平台上。一个关键的目标是展示这些大脑启发的计算方法如何在开放世界问题上取得卓越的性能，最根本的是在下一代光学计算机平台上部署。与当代深度网络相比，本文描述的基于大脑的网络基于异构元素和反馈介导的动态系统，并基于完全本地化的计算来运行，从而消除了对共享内存资源的需求。因此，它们学习速度很快，当部署在像英特尔Loihi这样的神经形态平台上时，它们表现出更快的速度，极大地减少了能量预算。重要的是，最先进的光子计算基板与神经形态计算架构直接兼容，这表明它们将成为这些分散的、受大脑启发的计算算法的引人注目的平台。该项目的智力价值在于开发、部署和基准测试一套已建立的分散的、受大脑启发的算法，这些算法旨在在各种平台（包括前沿光子计算基板）上不可预测的开放世界条件下成功进行感官识别。具体来说，这些算法将被扩展到包含高阶大脑启发电路特性，部署到现场使用的便携式设备平台上，并在光子基板上进行部署和测试，以展示这些计算平台的转换潜力。更广泛的影响包括两个pi继续致力于监督来自STEM中代表性不足的群体的学生在与这里提出的研究直接相关的项目中的本科研究经历，以及使用神经形态方法开发新一代智能设备的潜力。PI Cleland还打算将本应用程序中讨论的概念纳入他的高级本科神经表征研讨会课程的一个单元。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。

For some tasks, modern computers vastly outperform the human brain – for example, large-scale numerical calculations, or the precisely accurate recall of organized information. But for other important tasks, the brains of humans and other animals are far superior to any computing system that has been built, both in terms of what they can do and in terms of the startlingly low energy required. For example, people can recognize familiar individuals at a distance, not only by their facial features, but by gait or other subtleties of movement that they might not even be able to articulate. People and other animals rapidly acquire information from their environments, but also are able to intelligently apply that information under novel, unforeseen circumstances. The study of brain-inspired computing is devoted to learning fundamental new ways to think about how computers work with information, so that they can perform better on such weakly-defined, open-world problems. In parallel, advances in physics have produced new optical materials and methods that can perform computations very rapidly and with extraordinarily low energy costs – if the problems of interest can be structured in ways compatible with these brain-inspired computing techniques. This project seeks to develop a brain-inspired computing network that learns rapidly and solves a set of real-world identification tasks, and to deploy this network onto portable devices as well as custom testing platforms built with these advanced physical substrates. A key goal is to show how these brain-inspired computing methods can achieve superior performance on open-world problems, most radically so when deployed on next-generation optical computer platforms. In contrast to contemporary deep networks, the brain-inspired networks described in this proposal are based on heterogeneous elements and feedback-mediated dynamical systems, and operate based on fully localized computations that obviate the need for shared memory resources. Consequently, they learn rapidly, and when deployed on neuromorphic platforms such as Intel Loihi they exhibit increased speed and tremendously reduced energy budgets. Importantly, state of the art photonic computing substrates are directly compatible with neuromorphic computational architectures, suggesting that they will be compelling platforms for these decentralized, brain-inspired computing algorithms. The intellectual merit of this project is to develop, deploy, and benchmark an established set of decentralized, brain-inspired algorithms designed for successful sensory identification under unpredictable, open-world conditions on a range of platforms, including leading-edge photonic computational substrates. Specifically, the algorithms will be extended to incorporate higher-order brain-inspired circuit properties, deployed onto portable device platforms for use in the field, and also deployed and tested on photonic substrates to demonstrate the transformational potential of these computational platforms. Broader impacts include a continuing commitment by both PIs to supervising undergraduate research experiences for students from groups underrepresented in STEM on projects directly connected with the research proposed here, as well as the potential for development of a new generation of smart devices using neuromorphic methods. PI Cleland also intends to incorporate the concepts discussed in this application into a unit of his advanced undergraduate Neural Representations seminar course.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.