FET: Small: Smart Probabilistic Computation with Limited Resources

这个项目的总体目标是了解如何制造更智能、更紧凑、更环保的计算设备，以便有效地应对不确定性。目前许多基于计算机的机器，例如移动电话和先进的生物医学设备，依赖于远程数据中心(所谓的云)来提供类似大脑的计算服务，这些服务采用了人工智能，并且本质上是概率的。这些中心消耗了大量的高成本资源，如计算机硬件、计算时间和电能。以低成本、可靠和环保的方式将它们的功能直接嵌入智能手机等最终用户设备中，是一项具有广泛影响的关键研究挑战。实现这一点的一种方法是本项目中探索的随机计算(SC)方法。SC电路本质上是概率性的，比传统电路小几个数量级，容错性更强。因此，它们提供了新的方法来创建微型、低成本和可靠的系统。除了降低数据中心等大型系统的成本和需求外，SC最近被证明适合于资源极其有限的生物医学应用，如视网膜植入物，以恢复盲人的视力。该项目还将使研究生、本科生和代表性不足的学生能够在现代计算技术的前沿接受计算技术方面的培训，从而为未来的工作做出贡献。传统的计算机科学主要致力于非概率(确定性)方法，这些方法虽然比概率方法更容易理解，但不太适合上面提到的感兴趣的应用。随机计算是现有的最有前途的概率技术之一。它可以被简明地定义为用0‘S和1’S随机流编码的概率进行计算。它很容易用标准电子技术实现；例如，即使在输入数据噪声很大的情况下，单个逻辑门也可以执行乘法。该项目将研究供应链的理论基础及其在资源严重受限系统中的应用。需要解决的问题包括：SC精确度的计算基本限制是什么？如何利用SC独特的渐进精度和随机性特性？系统架构师如何有效地平衡准确性、性能、能源需求和总体系统成本？原型系统将通过使用现场可编程门阵列的数学分析、模拟和仿真来设计和评估。这些设计包括融入神经网络的设计，以及智能相机、医疗植入物和便携式自动语音识别系统。这项研究还有望为量子计算等其他概率技术提供有用的见解。这一奖项反映了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。

The overall goal of this project is to understand how to build smarter, more compact, and more environmentally friendly computing devices that can cope efficiently with uncertainty. Many current computer-based machines, for example, mobile phones and advanced biomedical devices, depend on remote data centers (the so-called cloud) to provide brain-like computational services that employ artificial intelligence and are probabilistic in nature. These centers consume enormous amounts of high-cost resources such as computer hardware, computing time, and electrical energy. Embedding their functions directly in low-cost, reliable and environmentally friendly ways in end-user devices like smartphones is a key research challenge with broad implications. One way of doing this is the stochastic computing (SC) approach explored in this project. The SC circuits are inherently probabilistic and are orders-of-magnitude smaller and more error-tolerant than conventional circuits. Thus, they provide novel ways to create tiny, low-cost and reliable systems. Besides its potential for reducing the cost of, and need for, large systems like data centers, SC has recently been shown suitable for extremely resource-restricted biomedical applications such as retinal implants to restore vision to the blind. The project will also enable graduate, undergraduate and underrepresented students to be trained in computing techniques at the forefront of modern computing technologies, thus contributing to the future workforce.Traditional computer science is mostly devoted to non-probabilistic (deterministic) methods that, although far easier to understand than probabilistic, are less well-suited to the applications of interest cited above. Stochastic computing is one of the most promising probabilistic techniques available. It can be concisely defined as computing with probabilities encoded in randomized streams of 0’s and 1’s. It is readily implementable with standard electronic technology; for example, a single logic gate can perform multiplication, even when the input data is very noisy. The project will investigate the theoretical underpinnings of SC and its application to severely resource-restricted systems. Among the questions to be addressed are: What are the computational fundamental limits on SC’s accuracy? How can SC’s unique progressive precision and randomness properties be exploited? How can system architects efficiently balance accuracy, performance, energy needs, and overall system cost? Prototype systems will be designed and evaluated via mathematical analysis, simulation and emulation using field programmable gate arrays. These include designs incorporating neural networks, as well as smart cameras, medical implants, and portable automatic speech-recognition systems. The research is also expected to provide useful insights into other probabilistic technologies such as quantum computing.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.