Talk by Suhas Diggavi (UCLA)
summary
Frame personalization as hierarchical empirical Bayes where clients sample parameters from an unknown population and plug‑in estimators yield James–Stein–style shrinkage that provably improves individual estimates for Gaussian targets. Unknown population variance requires a coarse hyperprior to stabilize alternating‑minimization, yielding the ADEPT method used for personalized autoencoders. For generation, a shared U‑Net backbone with client bias embeddings fits a two‑component mixture score model, lowering sample complexity and improving worst‑client performance; heterogeneous bandits show three regimes and central/shuffled DP gives near‑free privacy (polylog).
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:51Statistical framework for personalization
- 4:40James-Stein estimator benefits
- 7:34Hierarchical Bayes regularization
- 9:23ADEPT methodology
- 11:11Population models insights
- 14:25Diffusion models for personalization
- 18:26Shared representation in diffusion
- 22:00Online learning and privacy
- 25:00Collaboration in bandit problems
- 27:38Summary
related talks
41mKeynote: Personalization, Bandits, and Causal inference
46mLearning the Privacy-Utility Trade-off with Bayesian Optimization
47mPrivate targeting strategies
35mEstimating the Value of Personalization
01h 04mNonparametric Bayesian Methods: Models, Algorithms, and Applications II
01h 02mLarge-Scale Private Learning, Part I
39mModern Bayesian Experimental Design
01h 07mLarge-Scale Private Learning, Part II