Cyber-physical applications often analyze collected sensor data using machine learning algorithms. Many existing sensing systems lack intelligence about the target and naively generate large-scale data, making communication and computation significantly costly. In many cases, however, the data generated by sensors only contain useful information for a small portion of the sensor activity. For example, machine learning algorithms continuously process the visual sensors used for environmental/security monitoring to detect sensitive activities. Still, these sensors only carry out useful information for a short time. On the other hand, biological sensors intelligently generate orders of magnitude less amount of data. This project develops machine learning algorithms that provide real-time feedback to sensors to ensure they only generate data needed for learning purposes. The approach is expected to provide up to four orders of magnitude data reduction from sensors. The results from this research will broadly impact many sensors used in internet-of-things applications, including infrastructure, mobile devices, autonomous systems, robotics, and healthcare. The project will also support underrepresented minority students through synergistic outreach plans and educational activities, including programs for K-12 students, undergraduate research opportunities, and new course development.

The research approaches introduced in this project aim to make fundamental changes to sensing systems in order to make future sensors intelligent for a wide range of cyber-physical applications. First, this project will develop novel brain-inspired learning algorithms that can provide fast and real-time feedback to the sensing module to intelligently control the rate of data generation from sensors. This feedback also makes sensors aware of the target task, enabling situational awareness. Second, the project will develop a novel framework that tightly integrates with a sensing circuit and brain-inspired algorithms to dynamically control the sensor functionality in a close-loop manner. The proposed hardware platform exploits the robustness of learning algorithms to design near-sensor computing platforms that are highly approximate, parallel, and efficient. Finally, this project aims to evaluate the effectiveness of the framework on multiple large-scale systems. The prototype will be fully released under an established open-source library for public dissemination.