Why Diffusion Models Don't Memorize: The Role of Implicit Dynamical Regularization in Training
Raphaël Urfin (PhD Student, École Normale Supérieure - PSL)
| Quand ? |
Le 18/11/2025, de 13:00 à 14:00 |
|---|---|
| Où ? | M7 101 |
| Participants |
Raphaël Urfin |
| Ajouter un événement au calendrier |
vCal
iCal
|
Raphaël Urfin
Title: Why Diffusion Models Don't Memorize: The Role of Implicit Dynamical Regularization in Training
Abstract: Diffusion models have achieved remarkable success across a wide range of generative tasks. A key challenge is understanding the mechanisms that prevent them from memorizing training data and that allow generalization. In this talk, I will discuss the role of the training dynamics in the transition from generalization to memorization. Through extensive experiments and theoretical analysis, we identify two distinct timescales: an early time $\tau_\mathrm{gen}$ at which models begin to generate high-quality samples, and a later time $\tau_\mathrm{mem}$ beyond which memorization emerges. Crucially, we find that $\tau_\mathrm{mem}$ increases linearly with the training set size $n$, while $\tau_\mathrm{gen}$ remains constant. This creates a growing window of training times where models generalize effectively. These findings reveal a form of implicit dynamical regularization in the training dynamics, which allows to avoid memorization even in highly overparameterized settings. The talk is based on our recent work arXiv.2505.17638.
Website: arXiv.2505.17638
In Room M7 101, 1st floor, Monod campus, ENSL.