Summary of When Does Contrastive Visual Representation Learning Work? - CVPR 2022

00:00:00 - 00:00:00

This YouTube video discusses a research paper on contrastive visual representation learning, which examines how much unlabeled data is necessary for pre-training and how much label data is needed for linear classifier training or fine-tuning. The data quality and task granularity of the data used for pre-training is also examined. Finally, the paper discusses how self-supervised learning only gets close to fully supervised performance when lots of labels are available, but the gap between supervised and self-supervised learning remains quite large.

• 00:00:00 This YouTube video presents a research paper on contrastive visual representation learning, which examines how much unlabeled data is necessary for pre-training and how much label data is needed for linear classifier training or fine-tuning. The data quality and task granularity of the data used for pre-training is also examined. Finally, the paper discusses how self-supervised learning only gets close to fully supervised performance when lots of labels are available, but the gap between supervised and self-supervised learning remains quite large.