YINS Alumnae Seminar: Anup Rao (Adobe Research)
YINS Alumnae Seminar
“Machine Unlearning via Algorithmic Stability”
Speaker: Anup Rao
Research Scientist, Adobe Research
Formerly, beloved YINS graduate student advised by Daniel Spielman
Talk summary: We study the problem of machine unlearning, updating a trained machine learning model when part of the data is deleted. We identify a notion of algorithmic stability, Total Variation (TV) stability, and show why it is suitable for the goal of exact unlearning. For convex risk minimization problems, we design TV-stable algorithms based on noisy Stochastic Gradient Descent (SGD). Our key contribution is the design of corresponding efficient unlearning algorithms, and are based on constructing a (maximal) coupling of Markov chains for the noisy SGD procedure. To understand the trade-offs between accuracy and unlearning efficiency, we give upper and lower bounds on excess empirical and population risk of TV stable algorithms for convex risk minimization. This is based on joint work with Enayat Ullah, Tung Mai, Ryan Rossi and Raman Arora, and will be presented at COLT 2021.
To participate:
Join from PC, Mac, Linux, iOS or Android: https://yale.zoom.us/j/92015011527
Or Telephone:203-432-9666 (2-ZOOM if on-campus) or 646 568 7788
Meeting ID: 920 1501 1527
International numbers available: https://yale.zoom.us/u/acHGdkWp2U
.
.