Emergence of heavy tails in homogenised stochastic gradient descent
- 👤 Speaker: Martin Keller-ressel (Technische Universität Dresden)
- 📅 Date & Time: Thursday 25 April 2024, 15:15 - 16:00
- 📍 Venue: External
Abstract
We analyze a continuous diffusion approximation of SGD , called homogenized stochastic gradient descent, show that it behaves asymptotically heavy-tailed, and give explicit upper and lower bounds on its tail-index. We validate these bounds in numerical experiments and show that they are typically close approximations to the empirical tail-index of SGD iterates. In addition, their explicit form enables us to quantify the interplay between optimization hyperparameters and the tail-index. Our results show that also continuous diffusions, not only Lévy-driven SDEs, can accurately represent the emergence of heavy tails in SGD . In addition, our results suggest skew Student-t-distributions, not alpha-stable distributions, as surrogates of parameter distributions under SGD .
Series This talk is part of the Isaac Newton Institute Seminar Series series.
Included in Lists
- All CMS events
- bld31
- dh539
- External
- Featured lists
- INI info aggregator
- Isaac Newton Institute Seminar Series
- School of Physical Sciences
Note: Ex-directory lists are not shown.
![[Talks.cam]](/static/images/talkslogosmall.gif)

Martin Keller-ressel (Technische Universität Dresden)
Thursday 25 April 2024, 15:15-16:00