Name
Nima Anari
Date & Time
Wednesday, October 16, 2024, 2:30 PM - 3:00 PM
Speakers
Description
Speaker 5
Title
Parallel Sampling Via Auto-speculation
Abstract
I will talk about speeding up sampling via parallelism. We show surprising polynomial speedups (compared to sequential algorithms) are possible for sampling from *arbitrary* distributions in two settings: distributions on [q]^n accessed by conditional marginals (a la any-order autoregressive models), and distributions on R^n accessed by a denoising diffusion oracle.