CAREER: Hardware and Algorithmic Architectures for Analyzing Physically-complex Systems: embedding inference capabilities in ultra-low-power sensors

这项工作的科学目的是研究小型电子设备如何在面对日益严重的物理复杂性时发挥高水平的功能。复杂性可能源于感测系统中感兴趣的物理信号，或者源于电子器件本身中的非理想器件和算法行为，由于技术和系统缩放，这变得不可避免。例如，本研究的重点是分析通过低功耗医疗传感器提供的生理信号。虽然这样的信号是高度指示性的，但是提取有价值的医学信息需要基本生理过程的高阶模型，而实际上通常不存在易处理的分析模型。本研究还关注由于不可预测但不可避免的技术缺陷和变化而发生的硬件错误，这些错误导致正在处理的数据中出现高水平的错误。这些挑战是通过机器学习领域出现的算法方法来解决的，这些方法构建了用于从数据本身解释数据的模型。因此，可以利用小型传感器提供的大量数据作为广泛的知识库;但问题是，这些方法在计算能量、存储器要求、网络交互等方面没有得到低功耗电子设备的良好支持。本研究从机器学习框架中使用的内核计算开始，并研究内核公式，结构化的硬件体系结构和算法，以克服与应用信号和技术非理想性相关联的物理复杂性。通过硬件和软件的实验演示来研究这些原理。这项研究的更广泛的影响是使电子系统在关键应用中发挥更大的价值，并建立一个跨学科的教育计划，教导学生将计算机科学，低功耗电子学和临床应用的基础知识联系起来。虽然电子产品具有巨大的功能，但它对医疗保健等现实世界挑战的影响取决于与物理系统的高价值交互。该计划强调临床应用和合作，以了解电子设备在实现对大量患者人群的先发制人的医疗伤害检测和慢性病管理方面可以发挥的作用：这是当今方法不可行的。该计划还强调与半导体行业的互动，为先进的传感平台和硬件弹性的算法方法传递原理和架构;硬件错误已被业界确定为关键挑战之一，克服传统形式的设计余量的方法正在迫切追求。新的跨学科课程，学生项目和外展活动将使学生接触到外部合作者，并将推动一个教育计划，通过应用驱动的系统追求将多个领域的工程基础联系在一起，以克服医疗决策支持方面的关键挑战。

The scientific aim of this work is to study how small electronic devices can function at a high level in the face of increasingly-severe physical complexities. The complexities can originate from the physical signals of interest in a sensing system or from non-ideal device and algorithmic behaviors within the electronics itself, which are becoming unavoidable due to technology and system scaling. As an example, this study focuses on analyzing physiological signals that are available through low-power medical sensors. Though such signals are highly indicative, extracting medical information of value requires high-order models of the underlying physiological processes when in fact no tractable analytical models generally exist. This study also focuses on errors within the hardware that occur due to unpredictable but inevitable technological defects and variations, leading to high levels of errors in the data being processed. These challenges are approached through algorithmic methods emerging from the domain of machine learning that construct models for interpreting data from the data itself. The large amount of data that is available through small-scale sensors can thus be leveraged as an extensive knowledgebase; but the problem is that these methods are not well supported by low-power electronics, in terms of their computational energy, memory requirements, network interactions, etc. This research starts with the kernel computations used in machine-learning frameworks, and it investigates kernel formulations, structured hardware architectures, and algorithms to overcome the physical complexities associated with application signals and technological non-idealities. The principles are studied through hardware and software experimental demonstrations.The broader impact of this research is to enable greater value of electronic systems in critical applications and to establish an interdisciplinary educational program that teaches students to connect fundamentals from computer science, low-power electronics, and clinical applications. While electronics presents tremendous capabilities, its impact on real-world challenges such as in healthcare depends on high-value interactions with physical systems. This program emphasizes clinical applications and collaborations to understand the role that electronics can play in enabling preemptive medical harm detection and chronic-disease management over large patient populations: something that is infeasible with today's methods. This program also emphasizes interactions with the semiconductor industry, to transfer principles and architectures both for advanced sensing platforms and for algorithmic approaches to hardware resilience; with hardware errors having been identified by the industry as one of the critical challenges, methods that overcome the need for traditional forms of design margining are being urgently pursued. New interdisciplinary courses, student projects, and outreach activities will expose students to external collaborators and will drive an educational program that ties together engineering fundamentals from multiple domains through an application-driven pursuit of systems to overcome critical challenges in healthcare decision support.