Seminario Dottorato 2024/25 - Erik CHINELLATO (09.01.2025)

Deep neural networks (DNNs) have revolutionized numerous fields due to their powerful ability to learn complex representations. However, their black-box nature and lack of interpretability in architecture and weight design remain significant challenges. After an introductory segment on DNNs and backpropagation learning, this seminar introduces the Deep Unfolding method as a promising alternative, bridging the gap between data-driven learning and model-based optimization. By unrolling iterative optimization algorithms into structured neural network architectures, Deep Unfolding provides a principled approach to network design, enabling interpretability and theoretical insights into their operation. We will explore how this method leverages domain knowledge, achieves faster convergence, and enhances performance in resource-constrained scenarios. The session will highlight many wide-ranging practical applications of Deep Unfolding, covering audio source separation and recognition, image denoising and state estimation.