Active Reasoning in Sensor Networks

Carlos Guestrin

Abstract

Sensor networks consist of a collection of small low-cost devices that can sense and actuate in the environment, and communicate with each other through a wireless network. Recent advances in hardware and low-level software have made it possible for several real-world deployments to collect scientific and engineering data in complex, unstructured environments. Despite these recent advances, sensor networks pose unique challenges: Sensor nodes are highly constrained, with very limited energy, memory, computation and communication capabilities.Furthermore, in order to have significant impact in society, sensor nets must be able to solve more complex tasks than simple data collection, including probabilistic inference (e.g., for sensor calibration and target tracking), regression (e.g., for data modeling and contour finding), and optimization (e.g., for actuator control, decision-making and pattern classification). Machine learning techniques have recently demonstrated significant promise for solving such complex tasks, but are usually ill-suited for the limitations of sensor networks. Thus, it is not sufficient to simply apply ML algorithms directly to sensor nets, we require novel algorithms that directly address those limitations. In this talk, we present some of our recent advances, both theoretical and practical, in the marriage between machine learning and sensor networks. Interestingly, as we show, our approach has lead not only to novel contributions that significantly increase the capabilities of sensor nets, but also to new contributions to core ML problems.

Back to the Main Page