Collaborative Research:CNS Core: Small: Intermittent and Incremental Inference with Statistical Neural Network for Energy-Harvesting Powered Devices

合作研究：CNS 核心：小型：利用统计神经网络对能量收集供电设备进行间歇和增量推理

• 批准号: 2007274
• 负责人: Jingtong Hu
• 金额: $ 25万
• 依托单位: University of Pittsburgh
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-10-01 至 2023-09-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2007274&HistoricalAwards=false
• 关键词: Collaborative; Research; CNS; Core; Small

The maturation of energy-harvesting (EH) technology and the recent emergence of viable intermittent computing, which stores harvested energy in an energy storage and supports an episode of program execution, creates the opportunity to build sophisticated batteryless computing systems. This project aims to realize artificial intelligence (AI) in such batteryless devices. However, there are two main challenges: 1. most existing Deep Neural Networks (DNNs) are hard to fit in resource-constrained microcontrollers. 2. DNNs usually require multiple execution episodes to obtain one inference result and it may take indefinite amount of time due to the weak and unpredictable harvested power. To address these challenges, this project is developing multi-exit DNNs, which can output incrementally accurate inference results during each execution episode. Three tasks will be carried out to lay the technological foundation for intermittent incremental inference on EH-powered IoT devices. First, novel power trace aware compression, online pruning and adaptation algorithms will be developed to ensure efficient deployment of multi-exit DNNs on intermittently-powered devices. Second, new multi-exit statistical and incremental neural networks (MESI-NN) will be developed to further reduce the latency and improve the accuracy and energy efficiency. Third, new neural architecture search algorithms will be developed to automatically search the best MESI-NN architecture. This project will be evaluated with real system and applications such as image classification, keyword spotting, and activity recognition. Realizing AI in EH-powered batteryless devices can enable persistent, event-driven sensing capabilities in which the main device (e.g. a battery-draining camera) can remain off until awaken by the EH-powered device when it detects events of interest. The societal impact of the proposed research is to significantly extend the lifetime of sensors and devices deployed in remote areas, which will drastically benefit various consumer, business, scientific and national security applications. This project will expose students to related cutting-edge knowledge and hands-on research opportunities and elevate their competence and confidence in facing of today's highly competitive global job market. The education impact of the proposed research includes the integration of various education activities based on the resources available to the two PIs such as DAC System Design Contest; outreach for local K-12 students through Pitt’s Investing Now summer school and ND’s CS curriculum for K-12 students in Indiana; undergraduate research with emphasis on minority participation, and course integration of the research outcomes.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.

能量收集(EH)技术的成熟和最近可行的间歇计算的出现为构建复杂的无电池计算系统创造了机会。间歇计算将收集的能量存储在能量存储中并支持一段程序执行。该项目旨在在这种无电池设备中实现人工智能(AI)。然而，存在两个主要的挑战：1.现有的大多数深度神经网络(DNN)很难适用于资源受限的微控制器。2.DNN通常需要多次执行才能获得一个推理结果，并且由于获取的能力较弱且不可预测，可能需要不确定的时间。为了应对这些挑战，该项目正在开发多出口DNN，它可以在每个执行过程中输出增量准确的推理结果。将执行三项任务，为EH供电的物联网设备上的间歇增量推理奠定技术基础。首先，将开发新的功率跟踪感知压缩、在线修剪和自适应算法，以确保在间歇供电的设备上高效部署多出口DNN。其次，将开发新的多出口统计和增量神经网络(MESI-NN)，以进一步减少延迟，提高精度和能量效率。第三，将开发新的神经结构搜索算法来自动搜索最佳的MESI-NN结构。该项目将评估与真实的系统和应用程序，如图像分类，关键字识别，和活动识别。在EH供电的无电池设备中实现人工智能可以实现持久的、事件驱动的感应功能，其中主设备(例如电池耗尽的摄像头)可以保持关闭，直到被EH供电的设备唤醒时，它检测到感兴趣的事件。拟议研究的社会影响是显著延长部署在偏远地区的传感器和设备的寿命，这将极大地造福于各种消费、商业、科学和国家安全应用。该项目将使学生接触到相关的尖端知识和实践研究机会，并提高他们面对当今竞争激烈的全球就业市场的能力和信心。拟议研究的教育影响包括基于两个私人投资机构的现有资源整合各种教育活动，如DAC系统设计大赛；通过皮特的现在投资暑期学校和ND的印第安纳州K-12学生的CS课程向本地K-12学生推广；强调少数群体参与的本科生研究，以及研究成果的课程整合。该奖项反映了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。

项目成果

Intermittent Inference with Nonuniformly Compressed Multi-Exit Neural Network for Energy Harvesting Powered Devices

• DOI: 10.1109/dac18072.2020.9218526
• 发表时间: 2020-04
• 期刊: 2020 57th ACM/IEEE Design Automation Conference (DAC)
• 作者: Yawen Wu;Zhepeng Wang;Zhenge Jia;Yiyu Shi;J. Hu

Developing a Miniature Energy-Harvesting-Powered Edge Device with Multi-Exit Neural Network

• DOI: 10.1109/iscas51556.2021.9401799
• 发表时间: 2021-05
• 期刊: 2021 IEEE International Symposium on Circuits and Systems (ISCAS)
• 作者: Yuyang Li;Yawen Wu;Xincheng Zhang;Ehab A. Hamed;Jingtong Hu;Inhee Lee

Energy-Aware Adaptive Multi-Exit Neural Network Inference Implementation for a Millimeter-Scale Sensing System

毫米级传感系统的能量感知自适应多出口神经网络推理实现

• DOI: 10.1109/tvlsi.2022.3171308
• 发表时间: 2022
• 期刊: IEEE Transactions on Very Large Scale Integration (VLSI
• 作者: Li, Yuyang;Wu, Yawen;Zhang, Xincheng;Hu, Jingtong;Lee, Inhee
• 通讯作者: Lee, Inhee

Opportunistic Communication with Latency Guarantees for Intermittently-Powered Devices

针对间歇性供电设备的具有延迟保证的机会通信

• DOI: 10.23919/date54114.2022.9774732
• 发表时间: 2022
• 期刊: Automation & Test in Europe Conference & Exhibition (DATE
• 作者: Wardega, Kacper;Li, Wenchao;Kim, Hyoseung;Wu, Yawen;Jia, Zhenge;Hu, Jingtong
• 通讯作者: Hu, Jingtong

Lightweight Run-Time Working Memory Compression for Deployment of Deep Neural Networks on Resource-Constrained MCUs

• DOI: 10.1145/3394885.3439194
• 发表时间: 2021-01
• 期刊: 2021 26th Asia and South Pacific Design Automation Conference (ASP-DAC)
• 作者: Zhepeng Wang;Yawen Wu;Zhenge Jia;Yiyu Shi;J. Hu