Speaker |
Joshua David Robinson |
|---|---|
Affiliation |
MIT |
Date |
Friday, 07 October 2022 |
Time |
12:00-13:00 |
Location |
Function Space, UCL Centre for Artificial Intelligence, 1st Floor, 90 High Holborn, London WC1V 6BH |
Link |
https://ucl.zoom.us/j/97245943682 |
Event series |
DeepMind/ELLIS CSML Seminar Series |
Abstract |
Eigenvectors computed from data arise in various scenarios including principal component analysis, and matrix factorizations. Another key example is the eigenvectors of the graph Laplacian, which encode information about the structure of a graph or manifold. An important recent application of Laplacian eigenvector is to graph positional encodings, which have been used to develop more powerful graph architectures. However, eigenvectors have symmetries that should be respected by models taking eigenvector inputs; (i) sign flips, since if v is an eigenvector then so is -v; and (ii) more general basis symmetries, which occur in higher dimensional eigenspaces with infinitely many choices of basis eigenvectors. We introduce SignNet and BasisNet---new neural network architectures that are sign and basis invariant. We prove that our networks are universal, i.e., they can approximate any continuous function of eigenvectors with the desired invariances. Moreover, when used with Laplacian eigenvectors, our architectures are provably expressive for graph representation learning; they can approximate—and go beyond—any spectral graph convolution, and can compute spectral invariants that go beyond message passing neural networks. Experiments show the strength of our networks for molecular graph regression, learning expressive graph representations, and more. |
Biography |
Josh Robinson is a fifth year PhD student at MIT CSAIL working with Stefanie Jegelka and Suvrit Sra. His interests include self-supervised learning and models such as graph neural networks that operate on discrete structured data. |