Giovanni Ansaloni: how do you squeeze AI into power-constrained edge devices?

Edge intelligence has fostered a revolution in the way information is created, analyzed and accessed, fundamentally impacting multiple fields ranging from health monitoring to augmented reality. Key to this success is the optimization of application energy/run-time/application footprints, especially in light of the explosion in size of machine learning models vis-à-vis the tightly constrained computing and energy resources of edge architectures.

The resulting landscape calls for a careful tailoring of both “what” should be computed, as well as “how” such computation should take place. This course is an illustration of multiple strategies to address the ensuing challenges. Exploring the “what?” question, Dr. Ansaloni discusses opportunities in both features-based and ML-based approaches, harnessing domain-specific characteristics such as sparsity and robustness. As for the “how?” question, he describes different avenues towards the design of energy-minimal dedicated computing systems, ranging from smart SIMD architectures, to the use of in-and near-memory computing accelerators. The course has a strong focus on the interaction between those two viewpoints, also showcasing how open development hardware/software frameworks facilitate holistic explorations.