AreLarge-scaleSoftLabelsNecessaryforLarge-scale DatasetDistillation?

Neural Information Processing Systems 

In this paper, we first discover that the high within-class similarity in condensed datasets necessitates the use of large-scale soft labels.