EECS Special Seminar: Jason Altschuler, “Algorithmic Challenges in Optimization, Sampling, and Beyond”

Abstract: 
Optimization is the engine driving modern engineering and AI. Yet after a century of progress, fundamental questions remain open even about gradient descent in seemingly simple convex settings — highlighting the potential for untapped algorithmic opportunities in those foundational settings and beyond. This talk will present two such vignettes.

First, I’ll showcase an untapped opportunity for improving optimization algorithms. The key idea is that, surprisingly, combining individually bad optimization algorithms can create fast ones. I will illustrate this concretely through time-varying hedging of stepsizes, which has enabled us to dispel longstanding beliefs about the speed limit of gradient descent in convex optimization and min-max optimization. 

Second, I’ll discuss broader optimization settings that crucially involve probability and statistics — in particular connecting optimization, sampling, and differential privacy. I will argue that “shifted divergences” provide a powerful new approach towards unifying central problems in these fields. For concreteness I’ll focus on how this has enabled us to break longstanding complexity barriers of gradient-descent-type algorithms for log-concave sampling (the sampling analog of convex optimization).

Bio:
Jason Altschuler is an Assistant Professor at UPenn in the Department of Statistics and Data Science, and by courtesy also the Departments of Computer Science, Electrical Engineering, and Applied Mathematics. He received his undergraduate degree from Princeton and his PhD from MIT. He is the recipient of the Sloan Fellowship in Mathematics, the ICS Prize for the best papers at the interface of computer science and operations research, the AFOSR Young Investigator Program Award in Mathematical Optimization, the Mathematical Optimization Society’s Tucker Finalist Prize, the MIT Sprowls Dissertation Award, and Undergraduate Teaching Excellence Awards. His research interests lie at the interface of optimization, probability, and machine learning, with a focus on the design and analysis of efficient algorithms.  

Details

• Date: Monday, April 6
• Time: 10:00 am - 11:00 am
• Location: 32-D463 Star (CSAIL)

Host

• Greg Wornell