Cancelled: AI-Fun Seminar | Andrea Paudice: 'Tail Bounds for Non-Smooth Stochastic Convex Optimization Under Heavy-Tailed Noise'
| Dates: | 28 February 2024 |
|---|
| Times: | 10:30 - 11:30 |
|---|
| What is it: | Seminar |
|---|
| Organiser: | Faculty of Science and Engineering |
|---|
| Who is it for: | University staff, External researchers, Current University students |
|---|
| Speaker: | Andrea Paudice |
|---|
|
Unfortunately due to unforseen circumstances, we have had to cancel Andrea Paudice's talk. This will be rearranged for later in the year and will be advertised here.
The Manchester Centre for AI Fundamentals is hosting a series of seminars featuring expert researchers working in the fundamentals of AI.
On 28 February, Andrea Paudice joins us from the University of Milan and the Italian Institute of Technology Genoa.
Title
Tail Bounds for Non-Smooth Stochastic Convex Optimization Under Heavy-Tailed Noise
Abstract
In this talk, we revisit the non-smooth stochastic convex optimization framework under heavy-tailed noise. We show novel tail bounds on the convergence rate of stochastic mirror descent extending beyond the canonical Sub-Gaussian noise. In particular, under Sub-Weibull and polynomially tailed noise, we prove tail bounds for both the average and the last iterate. Some of these bounds feature an interesting "two-regime behavior" where the effect of the heavy tails vanishes. We also consider a more general noise model assuming only finite variance. In this case, we prove optimal tail bounds for a clipped-variant of stochastic gradient descent. Our results apply to a wide range of averages and extend to the case of kernel methods.
Bio
Andrea Paudice is a postdoctoral researcher at the University of Milan and an ELLIS member working on machine learning theory. Before that, he also held a position at the Italian Institute of Technology and the Imperial College London.
Andrea obtained his Ph.D. in computer science at the University of Milan "La Statale".
His current research interests are in stochastic convex optimization, generalization error bounds, and clustering with a focus on computationally and statistically efficient methods.
In the past, he has also worked in adversarial machine learning.
Speaker
Andrea Paudice
Organisation: University of Milan
Travel and Contact Information
Find event
1.218
University Place
Manchester
