Klinger et al.
Weighting schedules in score-based models determine which data features are learned and at what rate
By analyzing training dynamics at fixed noise levels, the authors show that only near the speciation time do both mode directions and relative weights become learnable, with weighting schedules controlling effective emphasis on that region.
Academic
Jérémie Klinger · Raphaël Urfin · Giulio Biroli · Marylou Gabrié
Laboratoire de Physique de l’École normale supérieure, ENS, Université PSL, CNRS, Sorbonne Université, Université Paris Cité
Research Digest··3 min read
The authors analyze the training dynamics of score-based generative models on multimodal Gaussian mixtures.
Why this paper
From Laboratoire de Physique de l’École normale supérieure, ENS, Université PSL, CNRS, Sorbonne Université, Université Paris Cité
In one line
Weighting schedules govern the rate at which each feature of multimodal data is acquired during score-based model training.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§