SHF: Small: Deep Neural Network Inference on Energy-Harvesting Devices

间歇供电的能量收集计算机是不需要电池或系留电源的复杂计算、传感和通信系统。 这些能量收集设备将成为下一代物联网（IoT）应用的基础，从可穿戴和植入式医疗设备到环境和大气监测，再到深空的微型ChipSat-scale卫星。 实现这些应用的价值需要能够频繁地在本地和自主地做出决策的智能设备（即，没有附近其他计算机的帮助）。例如，设备可能需要决定是否打开电池耗尽的摄像头来检测感兴趣的人，决定哪些嵌入在混凝土中的传感器能够收集关于老化桥梁的最有用的数据，以及决定何时以及从这些传感器发送多少数据回到云端。 近年来，使用深度神经网络的统计推断和机器学习已被证明是此类决策的最成功方法。机器学习是未来物联网设备的一个至关重要的功能，但当今资源受限的能量收集系统不支持深度神经网络推理所需的高强度计算。该项目构建了未来间歇性物联网设备所需的软件计算机系统和硬件计算机架构，以实现使用机器学习的自主智能决策。 该项目生产具有新颖算法的软件系统，使当今的能量收集物联网设备能够有效地做出智能决策。 然后，该项目将设计新颖的并行计算机架构，专门用于间歇输入功率的机器学习计算的有效操作。 这些架构将间歇性决策的效率进一步提高了数十倍或数百倍，从而实现了使用当今架构无法实现的新型智能物联网应用。这些软件和硬件组件的总和解决了间歇性系统上深度机器学习的存在问题，展示了其可行性，并实现了其对学术界，工业界和对社会重要的应用程序的好处，如国防，医疗保健和民用基础设施。 该项目通过与课程整合、对少数民族学生的辅导以及技术高中外展计划，为未来多元化的劳动力做出了贡献。该项目克服的关键挑战是使深度神经网络推理在资源受限的间歇性设备上可行。这项任务需要架构和软件支持，以容忍频繁的间歇性电源中断，并以数百微瓦的功率运行，而不是当今机器学习加速器所需的数十或数百毫瓦。 该项目开发近似的，间歇性的部分重新执行技术，以有效地容忍中断，而不需要不必要的检查点和恢复软件状态。 该项目开发了第一个安全的数据并行架构，将非易失性存储器与简单的计算元件阵列集成在一起。该项目包括一个软件和硬件原型的直接途径，并为未来的硅硬件实现的架构奠定了基础。该奖项反映了NSF的法定使命，并已被认为是值得通过评估使用基金会的智力价值和更广泛的影响审查标准的支持。

Intermittently powered, energy-harvesting computers are sophisticated computing, sensing, and communicating systems that do not need a battery or tethered power source. These energy-harvesting devices will form the foundation of the next generation of internet-of-things (IoT) applications, ranging from wearable and implantable medical devices, to environmental and atmospheric monitoring, to tiny ChipSat-scale satellites in deep space. Realizing the value of these applications requires intelligent devices that can frequently make decisions locally and autonomously (i.e., without help from other nearby computers). For example, a device may need to decide whether to turn on a battery-draining camera to detect a person of interest, to decide which sensors embedded in concrete to enable to collect the most useful data about an aging bridge, and to decide when and how much data to send from these sensors back to the cloud. In recent years, statistical inference and machine learning using deep neural networks has proven the most successful method for such decision-making. Machine learning is a crucially important feature for future IoT devices, but today's resource-constrained energy-harvesting systems do not support the high-intensity computations required by deep neural network inference. This project builds the software computer systems and hardware computer architectures required by future, intermittent IoT devices to enable autonomous, intelligent decision-making using machine learning. This project produces software systems with novel algorithms that enable today's energy-harvesting IoT devices to efficiently make intelligent decisions. This project will then design novel parallel computer architectures that are designed specifically for efficient operation of machine learning computations with intermittent input power. These architectures further increase the efficiency of intermittent decision-making by 10s or 100s of times, enabling a new class of intelligent IoT applications that are not possible using today's architectures. The sum of these software and hardware components addresses the existential question of deep machine learning on intermittent systems, demonstrating its viability and realizing its benefits to academia, industry, and in applications important to society, such as defense, healthcare, and civil infrastructure. This project contributes towards a diverse future workforce, through integration with course curricula, mentoring of students from under-represented minority groups, and technical high school outreach programs.The key challenge overcome by this project is to make deep neural network inference viable on a resource-constrained, intermittent device. This task requires architecture and software support to tolerate frequent, intermittent power interruptions and to operate with hundreds of microwatts of power instead of the tens or hundreds of milliwatts required by today's machine learning accelerators. This project develops approximate, intermittent partial re-execution techniques to efficiently tolerate interruptions without the need to unnecessarily checkpoint and restore software state. The project develops the first intermittence-safe data-parallel architecture, integrating non-volatile memory with an array of simple compute elements. The project includes an immediate path to software and hardware prototypes and lays the groundwork for a future silicon hardware implementation of the architecture.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.