Speaker: Enayat Ullah

\nAffiliation: Johns Hopkins Univ
ersity

Title: Machine unlearning via algorithmic stability

\nAbstract: We study the problem of machine unlearning\, and identify a not ion of algorithmic stability\, Total Variation (TV) stability\, which we a rgue\, is suitable for the goal of exact efficient 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\, which are based on constru cting a (maximal) coupling of Markov chains for the noisy SGD procedure. T o understand the trade-offs between accuracy and unlearning efficiency\, w e give upper and lower bounds on excess empirical and population risk of T V stable algorithms for convex risk minimization. Our techniques generaliz e to arbitrary non-convex functions\, and our algorithms are differentiall y private as well.

DTSTART;TZID=America/New_York:20210217T120000 DTEND;TZID=America/New_York:20210217T130000 LOCATION:https://wse.zoom.us/j/91450299380 SEQUENCE:0 SUMMARY:[Theory Seminar] Enayat Ullah URL:https://www.cs.jhu.edu/~mdinitz/theory/event/theory-seminar-enayat-ulla h/ X-COST-TYPE:free END:VEVENT END:VCALENDAR