Virtual Seminar Series
Benjamin Gess, Technische Universit盲t Berlin & Max Planck Institute for Mathematics in the Sciences.
- Time
- 2pm - 3pm
- Date
- 10 Jun 2026
- Duration
- 1 hour(s)
- Location
- Online
- Presenters
Benjamin Gess,
Technische Universität Berlin & Max Planck Institute for Mathematics in the Sciences.- Audience
- All Action members
- Registration Required
- No
For its Summer 2026 term, the STOCHASTICA SNIP seminar will be held online every two weeks starting on May 13. Talks of 30-45 minutes will be followed by 15-30 minutes of open discussion. Initially, seminars will be focused on introducing new problems and applications, to stimulate further discussions and ideas within individual working groups.
The ninth seminar will be given by Prof. Dr. Technische Universit盲t Berlin & Max Planck Institute for Mathematics in the Sciences), and chaired by (Action Science Communication Coordinator).
Catapults in Stochastic Gradient Descent in Machine Learning
We first recall the challenge of understanding the implicit bias of stochastic gradient descent (SGD) in machine learning. We then investigate the emergence of catapult dynamics in SGD. We provide proofs demonstrating that their behavior fundamentally differs from the deterministic case. The arguments are based on large deviations estimates, and are related to concepts in random dynamics. The presence of catapults is shown to cause the transition out of the lazy training regime and to promote convergence toward flatter minima.