Talk title: 

When Unlearning Fails: Probing the Robustness of Machine Unlearning

Abstract:

Machine unlearning (MU), the ability to selectively remove undesirable data or knowledge from trained models, has emerged as a promising tool for regulated and mission-critical domains. Yet, recent work reveals a critical weakness: much of the “forgotten” information persists and can be reactivated through lightweight interventions such as jailbreak prompts, quantization, fine-tuning, or replay attacks. In this talk, I will revisit these vulnerabilities and present principled approaches, inspired by robust ML research, to strengthen MU through optimization theory and optimizer design.

Bio: 

Sijia Liu is a Red Cedar Distinguished Associate Professor in the Department of Computer Science and Engineering at Michigan State University (MSU), and an Affiliated Professor at the MIT-IBM Watson AI Lab, IBM Research. His research centers on scalable and trustworthy AI, such as machine unlearning for vision and language models, scalable optimization for deep models, adversarial robustness, and data-model efficiency. He is a co-author of the textbook Introduction to Foundation Models (Springer, 2024). His honors include the NSF CAREER Award, the INNS Aharon Katzir Young Investigator Award, MSU’s Withrow Rising Scholar Award, and best paper honors at UAI (2022) and ICASSP (2017). He co-founded the New Frontiers in Adversarial Machine Learning Workshop series (ICML/NeurIPS 2021–2024) and has delivered tutorials on trustworthy and scalable ML at major conferences, including AAAI, NeurIPS, CVPR, ICCV, and ICASSP

Talk title: 

Do They See What We See? Mapping What Vision-Language Models Really Learn

Abstract:

The field of multimodal learning has witnessed significant progress in recent years, mainly enabled by advances in contrastive and autoregessive learning techniques. This talk aims to present some findings about how to understand the representations that are learned by those systems: First, I will discuss the abilities of Vision-Language models, namely with respect to spatial grounding and the problem of vision language alignment. Second, I will review recent works focusing on the explainability of those architectures. Finally, the talk will close with an analysis of biases in current evaluation benchmarks and how to address those. The talk will close with an outlook toward the challenges and future directions in multimodal learning to improve the efficiency and scalability of future systems.

Bio: 

Hilde Kuehne is a Professor for Computer Vision and Multimodal Learning at the Tübingen AI Center and an affiliated professor at the MIT–IBM Watson AI Lab.

Her research focuses on video understanding, learning without labels and the analysis of multimodal representations. She has created several highly cited datasets and works on large-scale video analysis, including the HMDB51 dataset, which received both the ICCV 2021 Helmholtz Prize and the IEEE PAMI Mark Everingham Prize.

She is GC for ICCV 2025 and currently serves as an Associate Editor for the IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI).