Dataset Distillation via Drifting
Synthetic training samples move toward real data and away from one another, without a classification objective.
The method works in a feature extractor’s latent space to approximate each class’s data distribution. The authors report better coverage of the real distribution than existing distillation methods, while retaining strong performance as classification training data. They also demonstrate model visualization and extensions beyond images, and propose distribution-based metrics as a complementary way to evaluate distilled datasets.
HF Daily Papers' note
The method works in a feature extractor’s latent space to approximate each class’s data distribution. The authors report better coverage of the real distribution than existing distillation methods, while retaining strong performance as classification training data. They also demonstrate model visualization and extensions beyond images, and propose distribution-based metrics as a complementary way to evaluate distilled datasets.
HF Daily Papers' note
score 5