Adversarial learning meets graphs (and why should you care?)
Location: Lecture Theatre B, Wolfson Building
A multitude of important real-world datasets (especially in biology) come together with some form of graph structure: social networks, citation networks, protein-protein interactions, brain connectome data, etc. Extending neural networks to be able to properly deal with this kind of data is therefore a very important direction for machine learning research, but one that has received comparatively rather low levels of attention until very recently.
With this recent surge in attention, many techniques commonly used when working with image data are breaking into the graph domain as well, and adversarial learning is no different. In this talk, I will attempt to motivate why this is a very important domain for adversarial learning experts to consider, both in terms of the plethora of open problems and the potentially deep consequences of solving them.
I will only assume knowledge of the essentials of convolutional neural networks and GANs. This talk will contain a crash course in graph-based neural networks, spanning the GCN (Kipf & Welling, ICLR 2017), MPNN (Gilmer et al., ICML 2017) and GAT (Veličković et al., ICLR 2018) models. Following that, I will outline three important research directions where these methods are fused with adversarial learning:
- Generative models on graphs, such as MolGAN (De Cao & Kipf, ICMLW 2018) and GCPN (You et al., NIPS 2018);
- Semi-supervised adversarial learning on graphs (GraphSGAN; Ding et al., CIKM 2018);
- Graph-based adversarial defence (PeerNets; Svoboda et al., ICLR 2019)