CAREER: Learning to Sense: Joint Learning of Task Oriented Cognitive Sensing with Data Driven Reconstruction and Inference

传感器是我们生活中不可或缺的一部分，帮助社会的交通、健康、安全和通信需求。传统的感测方法以固定的方式获取数据，而与数据被用于的任务无关。此外，数据处理流水线中的每个数据采集、重构和推理模块都是相互独立的，分别进行优化。这种方法导致了指数级的数据生成速度，导致了对当今传感系统的电力、存储、处理和通信需求的难以承受的需求。该项目的目标是通过开发一种自适应的、面向任务的和物理感知的数据到决策的管道，在数据驱动的学习框架中联合优化数据获取、重建和推理阶段，从而推进基于学习的传感和处理技术的科学。这项研究将为未来智能、自适应和资源高效的传感系统奠定基础，用于各种应用，包括生物医学成像、遥感、雷达和无线通信。该项目有三个相互关联的目标：(I)通过与正则逆问题的基本关系和受现有信号处理模型启发的可解释体系结构，开发基于学习的物理感知多维信号重构技术；(Ii)开发基于数学和学习的自适应和面向任务的测量设计方法，联合优化传感、重构和处理块，并展示其对现实世界问题的影响；(Iii)开发基于学习的数据决策框架，它直接从少量的学习测量中推断出可操作的信息(分类、参数估计)。计划的协同教育和外联活动的中心主题是提高K-12和大学生以及公众在传感系统、信号处理和机器学习方面的科学素养。由于传感技术处于信息感知和提取的前沿，对广泛的应用至关重要，因此该项目将对正在开发的旨在改善我们日常生活质量的传感技术产生重大影响，范围从相机的应用到生物医学成像，或者从智能家居技术到自动驾驶汽车。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。

Sensors are an indispensable part of our lives, assisting society’s transportation, health, safety, and communication needs. Conventional sensing approaches acquire data in a fixed fashion, independent of the task for which the data is being utilized. In addition, each of data acquisition, reconstruction and inference blocks in the data processing pipeline is independent of one another and optimized separately. This approach has led to exponential rates of data generation that creates an unbearable demand for power, storage, processing, and communication requirements in today’s sensing systems. The goal of this project is to advance the science of learning-based sensing and processing technologies by developing an adaptive, task-oriented and physics-aware data-to-decision pipeline, which jointly optimizes data acquisition, reconstruction, and inference stages in a data-driven learning framework. The proposed research will establish the foundations of future smart, adaptive, and resource-efficient sensing systems for a variety of applications, including biomedical imaging, remote sensing, radar, and wireless communications.This project has three interconnected objectives (i) Developing learning-based physics-aware multi-dimensional signal reconstruction techniques through foundational relations to regularized inverse problems and explainable architectures inspired from existing signal processing models, (ii) Developing mathematical and learning-based adaptive and task-oriented measurement design approaches with jointly optimized sensing, reconstruction and processing blocks, and demonstrate its impacts on real-world problems, (iii) Developing a learning-based data-to-decision framework, which infers actionable information (classification, parameter estimation) directly from low number of learned measurements. The central theme of planned synergistic educational and outreach activities is to increase the scientific literacy of both the K-12 and university students and the public regarding sensing systems, signal processing, and machine learning. Because sensing technologies are on the frontier of how information is perceived and extracted, and are essential to a wide range of applications, this project will have a high impact on sensing technologies being developed to improve the quality of our daily lives, ranging from applications of cameras to biomedical imaging, or from smart home technologies to autonomous vehicles.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.