Wed, Oct 16 at 12pm in 540AB Cory Hall
Speaker: Ali Rahimi
Title: Auto-Tuning Programs and Designs
Abstract:
I’ve built a lot computer vision systems, machine learning software, digital hardware, low level system software, and some analog hardware. Building these systems is fun, but evaluating them is tedious and rarely rewarding. A thorough performance evaluation can take a lot of time. A bad performance evaluation can derail an entire field for years.
I’ll present the Berkeley Auto-Tuner, a web service I built at Intel Berkeley Research in 2009. The service tunes the parameters of a workload against datasets, then generates thorough and easy-to-examine performance tables. At its simplest, the Auto-Tuner iterates your designs. Typically, it can improve the performce of a system by 10-30% for free. At its best, it’s a way to automatically explore a design space to generate best-of-breed systems. I’ll talk about these various uses and its guts, a distributed optimization engine that can recover the latent sparse structure of an optimization problem.
I’ll also highlight some more recent work, including a way to measure the distance between cell phones using the ambient sounds they hear, a way to print DNA cheaply, and a service that copies house keys online
from a single photo.
Bio:
Ali Rahimi copies house keys at Keys Duplicated. Before that, he was a research scientist at Intel Berkeley, Intel Seattle, and adjunct faculty in the Computer Science department at the University of Washington.