Title: How powerful are graph neural networks, and what can they reason about?
Abstract: We study the theoretical and algorithmic aspects of Graph Neural Networks (GNNs) - an effective framework for learning with graphs. In part (i), we characterize the representational power of GNNs and build a maximally powerful GNN with Graph Isomorphism Network(GIN) and Jumping Knowledge Network (JK-Net). In part (ii), we study the generalization of GNNs, with a focus on abstract reasoning tasks. Our theory is based on an algorithmic alignment framework, and draws connections with over-parameterized NN theory. We show GNNs well align with dynamic programming algorithms, and thus, can solve a broad range of reasoning problems.
This talk is based on the following papers:
What Can Neural Networks Reason About? https://arxiv.org/abs/1905.13211
How Powerful are Graph Neural Networks? https://arxiv.org/abs/1810.00826
Representation Learning on Graphs with Jumping Knowledge Networks
Bio: Keyulu Xu is a Ph.D. student at Massachusetts Institute of Technology (MIT) in the EECS department, where he works with Professor Stefanie Jegelka. He is a member of the Computer Science and AI Lab (CSAIL) and machine learning group. Keyulu has been a visiting researcher with Professor Ken-ichi Kawarabayashi at National Institute of Informatics (NII) since 2016. Keyulu is a also a research fellow with Hudson River Trading (HRT) AI Labs. His research interests span the theory and practice of algorithmic machine learning.
Public events of RIKEN Center for Advanced Intelligence Project (AIP)Join community