Unlearning in Generative Models
Pinak Mandal (Postdoc at the University of Sydney)
| Speaker |
Pinak Mandal (Postdoc at the University of Sydney)
|
|---|---|
| When |
Aug 27, 2025
from 04:00 PM to 05:00 PM |
| Where | LH-111, First Floor |
| Add event to calendar |
vCal
iCal
|
SEMINAR TALK
Title: Unlearning in Generative Models
Abstract: As generative models become increasingly powerful and pervasive, the ability to unlearn specific data, whether due to privacy concerns, legal requirements, correction of harmful content or unphysical outputs, has become critically important. Unlike conventional training, which accumulates data and reinforces knowledge, unlearning selectively removes the influence of specific data points without retraining from scratch. As model sizes increase, efficient unlearning algorithms become essential since the cost of training scales accordingly. In this presentation, we will begin with some intuitive baseline algorithms, examine their shortcomings, and then present a fast, new algorithm that utilizes loss orthogonalization to unlearn. Along the way, we will examine how the generated samples evolve after unlearning, and what becomes of the forgotten memories.
Speaker's Bio: I am currently a postdoc at the University of Sydney under Georg Gottwald and we work on problems in machine learning and dynamical systems. Before this I was a PhD student at the International Centre for Theoretical Sciences - Tata Institute of Fundamental Research under Amit Apte and Vishal Vasan.