CRII: SHF: On-chip Multi-Task Learning Analog Artificial Intelligence For Low-Cost Image-Based Environmental Monitoring

无线成像是一种重要的非破坏性环境监测工具，包括鸟类或濒危物种的栖息地监测，可以提供有关濒危物种的行为模式和分布的重要见解。对支持无线成像的成像器的关键要求是，它们应当从电池源汲取非常低的功率，使得电池持续很长时间，因为用于环境监测的成像器通常部署在无法访问有线电源的位置中。为了了解功率要求，无线成像仪的操作可以分为2个阶段- 1）图像采集和2）通过无线网络传输图像。CMOS成像器技术的最新进展显著降低了图像采集阶段的功耗。然而，图像传输仍然消耗比图像采集高几个数量级的能量，这将电池寿命限制为几周。该项目将产生超低功耗无线CMOS成像器，可以从标准电池电源运行几个月，而不仅仅是几个星期。虽然该项目的具体研究目标与无线成像器有关，但相同的原理可以扩展到为物联网（IoT）和可穿戴医疗保健设计高能效边缘设备。本项目所涉及的基础研究课题可能会吸引广泛的学生，并将被研究者用于涉及高中和本科生的外展计划，以激励他们在STEM领域攻读研究生课程。为了减少图像传输过程中的高能耗，本项目将利用人工智能（AI）通过以下双管齐下的方法来减少能量传输：a）压缩原始图像，以及B）仅在识别出感兴趣对象时传输图像。模拟电路设计技术将被用来实现人工智能算法的硬件在非常低的面积和能源成本。该项目有三个组成部分- 1）开发AI算法以降低传输功率，2）设计电路以在芯片上实现AI算法，以及3）验证项目目标。将开发一个多任务学习AI模型，该模型将在同一个神经网络内执行两个共享任务- a）压缩原始图像，B）识别感兴趣对象（自然栖息地中的目标动物物种），并仅传输感兴趣对象的压缩图像。为了抑制与模拟设计相关的非理想性，将使用硬件-软件协同设计方法，其中物理晶体管模型被纳入离线AI模型训练阶段，以极大地抑制软件训练和硬件实现结果之间的偏差。该项目将产生嵌入AI模型的CMOS芯片，并将使用来自公开数据集（如CIFAR-100）的野生动物图像进行测试。该奖项反映了NSF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。

Wireless imaging is an important tool for non-disruptive environmental monitoring, including habitat monitoring of birds or endangered species which can provide important insight about behavioral pattern and distribution of endangered species. The key requirement for imagers supporting wireless imaging is that they should draw very low power from battery source such that the battery lasts a long time, since the imagers used for environmental monitoring are typically deployed in locations without access to wired power source. In order to understand the power requirement, operation of wireless imager can be divided into 2 phases – 1) image acquisition and 2) transmission of image over the wireless network. Recent advances in CMOS imager techniques has significantly reduced power consumption during image acquisition phase. However, image transmission still consumes several orders of magnitude higher energy than image acquisition which limits battery life to few weeks. The project will result in ultra-low power wireless CMOS imagers that can run from standard battery source for several months instead of just weeks. While the specific research aims of this project relate to wireless imagers, the same principles can be extended to design high energy-efficiency edge devices for internet-of-things (IoT) and wearable healthcare. The fundamental research topics addressed in this project is likely to appeal to broad set of students and will be leveraged by the investigator for outreach programs involving high school and undergraduate students to motivate them to pursue graduate studies in STEM fields.To reduce high energy consumed during image transmission, this project will leverage artificial intelligence (AI) to reduce energy transmission by adopting the following two-pronged approach: a) compress raw images, and b) transmit images only upon identification of object-of-interest. Analog circuit design techniques will be used to implement the AI algorithms in hardware at very low area and energy cost. The project has three components – 1) development of AI algorithms to reduce transmission power, 2) design of circuits to implement the AI algorithms on-chip, and 3) validation of the project aims. A multi-task learning AI model will be developed which will perform two shared tasks within the same neural network – a) compress the raw image, b) identify object-of-interest (target animal species in natural habitat) and only transmit compressed image of the object-of-interest. To suppress non-idealities associated with analog design, a hardware-software co-design methodology will be used in which physical transistor models are incorporated into the offline AI model training phase to greatly suppress deviations between software training and hardware implementation results. The project will result in a CMOS chip with the AI model embedded and will be tested with images of wildlife from publicly available dataset (such as CIFAR-100).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.