----------------------------------------------------------------------------------------------------

     Department of Systems Engineering and Engineering Management

                    The Chinese University of Hong Kong

----------------------------------------------------------------------------------------------------

Date: Friday, December 04, 2026, 4:30pm to 5:30pm HKT

Venue: ERB909, The Chinese University of Hong Kong

Title:  Graph-Signal Co-Evolution: An Interpretable Foundation for Lightweight Transformers

Speaker: Prof. Gene Cheung, York University

Abstract:

Recent advances in AI have largely relied on scaling black-box models to enormous sizes. We argue that a complementary path forward is to develop parsimonious learning systems whose internal representations and computations remain interpretable to humans. In this talk, we present a graph-signal co-evolution framework as an alternative foundation for transformer-like neural networks. At each layer, a graph-learning module constructs a sparse signal-dependent similarity graph, while a graph-filtering module updates signal representations using the learned graph as a structural prior. We show that graph learning with normalization can be interpreted as a variant of self-attention, establishing a direct connection between attention and adaptive graph construction. Moreover, ideal low-pass graph filtering corresponds to low-frequency Laplacian eigenspace projection, and thus repeated learned graph filtering can be interpreted as recursive structured subspace learning. This co-evolutionary perspective leads naturally to interpretable and lightweight architectures requiring dramatically fewer parameters than conventional transformers. Experiments in image restoration, traffic prediction, and EEG classification demonstrate state-of-the-art performance and substantial gains in parameter efficiency.

Biography:

Gene Cheung is Professor and York Research Chair at York University, Toronto, Canada. He received the M.S. and Ph.D. degrees in Electrical Engineering and Computer Science from the University of California, Berkeley, in 1998 and 2000, respectively. From 2000 to 2009, he was a Senior Researcher at Hewlett-Packard Laboratories Japan, Tokyo. From 2009 to 2018, he was Assistant and subsequently Associate Professor at the National Institute of Informatics (NII), Tokyo, Japan. His research interests span graph signal processing, 3D imaging, and signal-centric approaches to machine learning. He has served on the editorial boards of several leading IEEE journals and is currently Senior Area Editor for IEEE Transactions on Signal Processing. He is a co-author of multiple award-winning papers and a recipient of the Canadian NSERC Discovery Accelerator Supplement (DAS) in 2019. He is an IEEE Fellow.   

Everyone is welcome to attend the talk!

SEEM-5201 Website: http://seminar.se.cuhk.edu.hk