Optimization unplugged 2025

Non-convex landscapes, dynamics and learning, August 25–27 in Lausanne

We aim to bring together experts and young researchers to discuss aspects of non-convex optimization that are related yet usually call for different mathematical tools: the study of landscapes, the analysis of algorithm dynamics, and the understanding of generalization in learning.

The schedule is kept light to favor interactions. We encourage blackboard talks where appropriate—the venue is optimized for that!

—The organizers: Nicolas Boumal, Lénaïc Chizat, Nicolas Flammarion and Andrew McRae.

Monday Aug. 25

9:00 Welcome
9:10 Irène Waldspurger Second-order optimization for Burer-Monteiro factorizations
9:50 Aymeric Dieuleveut Does heavy ball accelerate on smooth and strongly convex functions, beyond quadratics?
10:30 Break
11:00 Conversation starters Think “spotlights” but to entice chats rather than convey facts.
12:00 Lunch We’ll distribute vouchers for a campus restaurant.
14:00 Alex Damian Understanding optimization in deep learning with central flows
14:40 Conversation starters
15:40 Break
16:10 Vianney Perchet Last iterate convergence for zero-sum games
16:50 Drinks

Tuesday Aug. 26

9:00 Julien Mairal MAP estimation with denoisers: convergence rates and guarantees
9:40 Shuyang Ling Local geometry determines global landscape for synchronization in \(O(d)\)
10:20 Break
10:50 Clément Royer Nonconvex landscapes and the computation of matrix square roots
11:30 Radu Dragomir A theoretical explanation of grokking in the small regularization limit
12:10 Lunch
14:00 Vaidehi Srinivas Guarantees for alternating least squares (ALS) in overparameterized tensor decomposition
14:40 Andre Wibisono Mixing time of the proximal sampler in relative Fisher information via strong data processing inequality
15:20 Break
15:50 Quentin Berthet Implicit diffusion: efficient optimization through stochastic sampling
16:30 Free time
18:00 Still here? Plan ahead: the boat won’t wait! (Take metro M1 to Flon then M2 to Ouchy.)
18:50 Dinner Boat “Henry Dunant” departs from CGN port in Lausanne (Ouchy).

Wednesday Aug. 27

9:00 Aurélien Lucchi Rethinking assumptions in neural networks optimization
9:40 Pierre Marion Logistic regression with large step sizes
10:20 Break
10:50 John Duchi Geometry, computation and optimality in stochastic optimization
11:30 Lunch bags + hike We go to Villeneuve and walk (on pavement) to Montreux (~1h30). From there, the world is your oyster!

EPFL stands by a lake next to the Alps. You can see the lake and the Mont Blanc from the Bernoulli Center’s balconies. ©Istock/VogelSP

This event is sponsored by the Bernoulli Center and by the Swiss State Secretariat for Education, Research and Innovation (SERI) under contract number MB22.00027.