核心概念
Generative models enhance self-supervised learning by producing diverse and semantically consistent image augmentations.
統計
"Our empirical study with ICGAN and Stable Diffusion models demonstrates the effectiveness of the generative transformations for self-supervised representation learning."
"The model which uses generative transformations with Stable Diffusion outperforms the baseline by 2.1% in Top-1 accuracy."
"In this experiment on some datasets, the ICGAN transformation performs similarly or better than the baseline."
引用
"Our new transformation is based on conditional generative models."
"Generative models offer a richer set of data for self-supervised learning."
"Generative transformations preserve semantics and produce realistic images for SSL."