Implicit Bias in Gradient Flow Learning: Results, Open Questions, and Emerging Dynamics
Loucas Pillaud-Vivien (assistant professor at Ecole des Ponts ParisTech, CERMICS)
| When |
Jan 27, 2026
from 01:00 to 02:00 |
|---|---|
| Where | M7 101 |
| Contact Name | Cédric Gerbelot |
| Attendees |
Loucas Pillaud-Vivien |
| Add event to calendar |
vCal
iCal
|
Loucas Pillaud-Vivien
Title: Implicit Bias in Gradient Flow Learning: Results, Open Questions, and Emerging Dynamics
Abstract:
In this talk, I will present several results on the estimator selected by gradient flow in overparameterized learning problems. This phenomenon, known as implicit bias, plays a central role in modern machine learning, as it governs generalization in the absence of explicit regularization.
Beyond classical settings where the implicit bias is now well understood, I will discuss a range of open questions arising in more complex models and learning regimes. In particular, I will focus on dynamical aspects of gradient-based learning, including saddle-to-saddle dynamics and connections with incremental learning, which reveal a richer and still largely unexplored structure of these systems.
Website: https://thebiglouloup.github.io/loucaspillaudvivien/
In Room M7 101, 1st floor, Monod campus, ENSL.