Adversarial Training for Good

Chen Zhu (University of Maryland)

Chen Zhu is a PhD candidate at University of Maryland, College Park. His adviser is Prof. Tom Goldstein. His research focuses on developing better algorithms to enhance the accuracy, efficiency and robustness of neural networks for applications in Computer Vision and Natural Language Processing. Before coming to UMD, he obtained a master’s degree from ShanghaiTech University, where he worked with Prof. Kewei Tu and Prof. Yi Ma, and a bachelor’s degree from Beihang University. Meanwhile, he has worked as a research intern at Google, Nvidia, Microsoft, Baidu and Intel.

Short Abstract: Adversarial training is an effective technique to improve the robustness of neural networks, but it was believed to cause worse generalization. In this talk, we show that adversarial training can improve the generalization in several domains, including language understanding, vision-and-language, and graphs. We introduce an enhanced adversarial training algorithm that promotes higher invariance in the embedding space of pre-trained Transformer models and achieved SOTA results for language understanding. We also provide the first known effort on large-scale adversarial training for vision-and-language. Lastly, we show adversarial training improves generalization of GCNs for node classification, link prediction and graph classification.