Micro AI: When Intelligence Moves to the Low Power Sensors
Tinoosh Mohsenin, UMBC
Abstract: Artificial intelligence is being used in a variety of edge-computing devices such as biomedical sensors, wearables and autonomous systems. Processing these sensor-level machine learning tasks come at the cost of high computational complexity and memory storage which is overwhelming for these light weight and battery constrained devices. Equally important is the need for designing smarter AI systems that can reason over in the face of a highly variable and unpredictable world. This talk overviews some research solutions that enable performing data analytics from a variety of multimodal sensors in real time while consuming low power. I will also talk about adding reasoning in these systems to improve acting and learning performance. Combining these solutions will bring exciting opportunities for future micro AI processors
Abstract: The early designers of the Internet fostered tremendous innovation by leaving much of the network’s functionality to the programmable computers at its periphery. Unfortunately, the *inside* of the network has been much harder to change. Yet, changing the network is important to make the Internet more reliable, secure, performant, and cost-effective. The networking research community has struggled for many years to make networks more programmable. What has worked, and what hasn't, and what lessons have we learned along the way? This talk offers my perspective on these questions, through a 25-year retrospective of research on programmable networks, focusing on my own research experiences as well as reflections on major trends in the field. The talk advocates a sort of “ambitious pragmatism” that approaches an ambitious long-term goal (a programmable network infrastructure) through smaller, pragmatic steps while keeping an eye on the prize..