Luca Pezzarossa

Technical University of Denmark, Denmark

Embedded AI: Small Models Are Not Simple

Abstract

Embedded AI is often presented as the challenge of fitting small neural networks onto constrained devices. But small models are not simple: making them useful in practice requires careful design across models, quantization, compilation, runtime execution, and embedded hardware architectures. This talk presents our group’s activities at DTU Compute in Embedded AI, covering adaptive inference, efficient audio processing, quantization, compilation, and on-device learning. The talk will frame Embedded AI as a full-stack systems challenge and also discuss the need for architectural support for mixed precision, dynamic execution, and efficient adaptation at the edge.

Biography

Luca Pezzarossa is an Associate Professor in the Embedded Systems Engineering section at DTU Compute, Technical University of Denmark. His research focuses on embedded systems, embedded AI, real-time and reconfigurable architectures, and programmable microfluidic biochips. He received his Ph.D. from DTU, working on reconfigurable computing for real-time embedded systems, and his M.Sc. and B.Sc. from Marche Polytechnic University, Italy. His current work explores efficient and predictable embedded computing, including embedded AI, chip design, real-time architectures, and programmable microfluidic systems.