From neural fields to physically *consistent* machine learning
By Diego Di Carlo

Diego Di Carlo will give a talk on neural fields and physically consistent machine learning.

Abstract

This seminar explores neural fields as a framework for continuous signal representation, with a focus on their application to physics-based problems. These models enable infinite resolution processing and straightforward multimodal design, addressing the limitations of traditional discrete representations. The seminar covers key architectures and their extension into Physics-Informed Neural Networks (PINNs), demonstrating their use in audio applications such as spatial filtering for speech enhancement, continuous sound source localisation, and sound field reconstruction.

Biography

Diego Di Carlo received his Master’s in Computer Engineering from the University of Padova in 2017 and a Ph.D. in Audio Signal Processing from Université de Rennes 1 and INRIA in 2020. After a one-year postdoc at Université de Rennes 2, focusing on Physics-Informed Machine Learning, he joined RIKEN AIP in Japan in 2022. His research centers on spatial audio processing, particularly applying neural generative models with physical constraints with application in augmented and virtual reality.