Abstract:

Artificial intelligence (AI) especially deep learning models have been widely deployed in many safety-critical applications such as face recognition and autonomous driving, while its trustworthiness is still a big concern. In this talk, we take a look at trustworthy deep learning from a reliability and robustness perspective, with an emphasis on generalization and robustness of deep neural networks subject to new and adversarial data. In particular, we consider generalization (to new data) and robustness (to adversarial data) from a randomized weight perspective, where model parameters are treated as random variables. Through the analysis of (robust) generalization error bounds, we propose several new regularization techniques to enhance generalization and adversarial robustness, and verify the reliability and robustness performance over different image classification datasets.

Speaker Bio:

Xinping Yi is currently a Professor at the National Mobile Communications Research Laboratory, Southeast University, Nanjing, China. He received his Ph.D. degree in Electronics and Communications from Telecom ParisTech, Paris, France. Before coming back to China, he has been a Lecturer at University of Liverpool, United Kingdom, a postdoctoral researcher at Technical University of Berlin, Germany, and a research assistant at EURECOM, France. His main research interests include information theory, machine learning, graph neural networks, and their applications in wireless communications and artificial intelligence.