Talk by Prof. Mayank Vatsa
Title: Ctrl+Z for GenAI: How Models Unlearn and Learn to Understand “Not”
Abstract: Generative AI has advanced rapidly in content creation, but remains less capable when it comes to deletion and handling absence. This talk examines that challenge of control through two complementary lenses: machine unlearning (teaching models to forget) and multimodal negation (teaching models to understand “no”). I begin with the notion of an AI “Ctrl+Z”: methods to selectively remove outdated, sensitive, or harmful knowledge without full retraining, spanning gradient-based unlearning and targeted model editing. I then turn to how models reason about absence and negation across text, vision, and audio, and why current systems often fail on prompts such as “write a story without romance” or “a room with no furniture.” These failures expose deeper limitations in compositional reasoning and constraint satisfaction. I will connect these issues to emerging privacy and regulatory requirements, as well as the broader goal of reliably following nuanced, constraint-heavy instructions. The talk concludes with an overview of current benchmarks, empirical insights from our recent work, and key open challenges in scaling these safety mechanisms to next-generation foundation models.
Bio: Mayank Vatsa obtained his MS and PhD degrees in Computer Science from West Virginia University, USA, and is currently a Professor at IIT Jodhpur, India. He is a Fellow of the IEEE and IAPR, and the recipient of the prestigious Swarnajayanti Fellowship from the Government of India and the IEEE Meritorious Service Award (2023). He served as the Vice President (Publications) of the IEEE Biometrics Council (2015–2018), where he led the establishment of the IEEE Transactions on Biometrics, Behavior, and Identity Science. He has served as an Area/Associate Editor for Pattern Recognition and Information Fusion, and as a General Co-Chair or PC Co-Chair for major biometrics conferences including IJCB, FG, AVSS, and ICB. Additionally, he has served as an Area Chair for premier vision/AI conferences such as CVPR, ICCV, ECCV, and AAAI. Mayank has contributed to different Indian government initiatives, including UIDAI (Aadhaar), the design of biometric standards for e-Gov applications, Responsible AI, and DigiYatra. He was instrumental in the formation of the Technology Innovation Hub (TIH) for Computer Vision and AR/VR at IIT Jodhpur. Currently, he leads the NCERT efforts to introduce AI to Class XI and XII curriculum and serves as a member of the Aadhaar Vision 2032 Expert Committee.
- 27 - 27 November,2025
- 06:00 pm - 07:00 pm
- SIT 001