Workshop Description

Graph representation learning and its applications have gained significant attention in recent years. Notably, Graph Signal Processing (GSP) and Graph Neural Networks (GNNs) have been extensively studied. GSP extends the concepts of classical digital signal processing to signals supported on graphs. Similarly, GNNs extend the concepts of Convolutional Neural Networks (CNNs) to non-Euclidean data modeled as graphs. GSP and GNNs have numerous applications such as semi-supervised learning, point cloud semantic segmentation, prediction of individual relations in social networks, image, and video processing. Early GSP researchers explored low-dimensional representations of high-dimensional data via spectral graph theory, i.e., mathematical analysis of eigen-structures of the adjacency and graph Laplacian matrices. Researchers first developed algorithms for low-level tasks such as signal compression, wavelet decomposition, filter banks on graphs, regression, and denoising, motivated by data collected from distributed sensor networks. Soon, researchers widened their scope and studied GSP techniques for image applications (image filtering, segmentation) and computer graphics. More recently, GSP tools were extended to video processing tasks such as moving object segmentation, demonstrating its potential in a wide range of computer vision problems.

GNNs have shown great potential in computer vision applications such as point cloud semantic segmentation, video understanding, and event-based vision. However, designing GSP or GNN algorithms for specific computer vision tasks has several practical challenges such as spatio-temporal constraints, time-varying models, and real-time implementations. Indeed, the computational complexity of many existing GSP/GNN algorithms at present for very large graphs is currently one limitation. In semi-supervised learning, GSP-based classifiers provide clear interpretations from a graph spectral perspective when propagating label information from known to unknown nodes. However, centralized graph spectral algorithms are slow and no fast-distributed graph labeling algorithms are known to perform well. In that sense, research is required in the development of fast GSP/GNN tools to be competitive against well-established deep learning methods like CNNs.