On Robustness and Generative Modeling скачать в хорошем качестве

On Robustness and Generative Modeling

Date Presented: 6/28/2024 Speaker: Iacopo Masi Abstract: The bloom of AI capabilities and their practical implication in our lives raised concerns regarding AI’s robustness to an adversary. A way to improve the robustness of the prediction is adversarial training (AT), which trains the predictor on adversarial data. Although AT is mainly associated with discriminative models, in this talk, I will show how we can shed light on some mysterious behaviors using generative modeling. By reinterpreting a robust discriminative classifier as an Energy-based Model (EBM), we offer a new take on the dynamics of adversarial training. On the ground of our thorough analysis, we present new theoretical and practical results that show how interpreting AT energy dynamics unlocks a better understanding of (i) robust overfitting, (ii) distinct features of different AT methods, such as SAT and TRADES (iii) their generative capabilities; also offering a simple process to lift these capabilities without training for generative modeling. In the last part of the talk, I will discuss how we can use off-the-shelf robust classifiers to help generative modeling by inverting them. Speaker's Bio: Dr. Iacopo Masi is an Associate Professor in the Computer Science Department at Sapienza, University of Rome. He is also the Principal Investigator and founder of the OmnAI Lab. Until August 2022, he held the position of Adjunct Research Assistant Professor in the Computer Science Department at the University of Southern California (USC). Previously, Dr. Masi was a Research Assistant Professor and Research Computer Scientist at the USC Information Sciences Institute (ISI). He has served as an Area Chair for several computer vision conferences (WACVs, ICCV'21, ECCV'22, CVPR'24) and is an Associate Editor for The Visual Computer - International Journal of Computer Graphics. Additionally, he organized an International Workshop on Human Identification at ICCV'17, co-organized the Unlearning and Model Editing (U&Me) workshop at ECCV'24, and is a general chair for ICIAP'25. In 2018, Dr. Masi was honored with the prestigious Rita Levi Montalcini Award by the Italian government. His primary research interests revolve around the intersection of machine learning, computer vision, and biometrics. Currently, he is exploring various interconnected lines of research, including adversarial robustness, proactive defense against image manipulation, inverse problems, and generative AI in both the vision and NLP domains.