Machine Unlearning: An Introduction
Date:
This talk gave an introduction to machine unlearning for the purposes of removing biases, confusion, and protecting user privacy. It focused on the premise of unlearning, the difference between exact and approximate unlearning, and methods which approach unlearning. Additionally, I discussed major open problems in unlearning and the challenges in the field. These focused on problems regarding data, metrics and evaluation, and the need for a unified ‘goal’ for unlearning. I ended the talk by doing a deep dive into how these problems are exacerbated in the cases of Large Language Models, and recent efforts (TOFU and WMDP) to improve the quality of work in the field.