Alexandre Vérine - Quality and Diversity in generative models through the lens of f-divergences.
- Date: 24 novembre 2025 à 13h
- Salle: 65-66 304
Generative modeling is a fundamental tool in machine learning for synthesizing realistic samples from complex data distributions, with models such as GANs, Normalizing Flows, Diffusion Models for visual data, or
LLMs for textual data. However, balancing sample quality and diversity remains a persistent challenge, making precision (sample quality) and recall (sample diversity) essential evaluation metrics. This talk
presents a unified view of precision and recall within the broader framework of f-divergences, providing a cohesive metric system alongside a tractable estimation method that enables the optimization of their
trade-off during model training across modalities. Building on this framework, we introduce a dedicated study of precision and recall within specific subgroups of interest inside the original data distribution,
enabling the detection of subgroup-level performance disparities. This approach provides a principled way to assess fairness in generative models.