Machine Learning for Network Data Workshop
Abstract
The goal of this proposal is to fund a workshop to discuss our incipient understanding of how we can take advantage of the underlying irregular structure of a graph signal to produce transformations that take advantage of the signal s internal symmetries. In particular, we will cover advances on graph signal processing, graph neural networks, and graph scattering transforms. Graph Signal Processing. Graph signal processing is an emerging field whose goal is to extend classical signal processing tools to signals supported on graphs. Central to this effort are generalizations of Fourier transforms, wavelets, and convolutions to irregular domains. Graph Signal Processing has had remarkable success in classical processing tasks such as filtering, systems identification, and sampling. How does the field need to evolve to address the higher cognitive questions that are typical of machine learning? Graph neural networks are layered architectures akin to Convolutional Neural Networks that use graph convolutions in lieu of standard convolutions. There are several architectures that have been proposed in the literature that build on different possible ways in which convolutions can be generalized. Naturally, graph neural networks lag behind convolutional neural networks in the richness of their architectures and their application domains. How can we expand the reach of graph neural networks into new application domains? How can we expand the range of architectures to process signals supported on graphs? Why are convolutional neural networks useful information processing architectures? Scattering transforms provide an answer to this question in the form of the notion of Lipschitz stability with respect to deformations. Graph scattering transforms have been introduced as an attempt to answer the analogous questions for graph neural networks. While preliminary stability results have been established, our understanding of stability properties of graph neural networks remains limited. How can we further our understanding of stability properties of graph neural networks? What other properties of graph neural networks are important to understand their behavior?
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 19, 2019
- Source ID
- W911NF1910096
Entities
People
- Alejandro Ribeiro
Organizations
- Army Contracting Command
- United States Army
- University of Pennsylvania