No. If I understand correctly, what they are doing is generating "noise" samples with statistics similar to natural images. Then they use unsupervised contrasti

No. If I understand correctly, what they are doing is generating "noise" samples with statistics similar to natural images. Then they use unsupervised contrastive learning to create representations of these noise images. This network is then employed in some classification tasks and it does well under a specific training mode (linear classifier training in the final layer only). The details of the evaluation aside, what is being shown is that a network trained to generate representation based on these artificially generated images (noise) can encode a good prior for most vision tasks thereby potentially reducing the need for a very large number of real training images. edited: Typo,clarity