Can anybody explain to me why does self-supervised representation learning on images using Siamese neural networks (such as SimpleSiam (https://arxiv.org/abs/2011.10566), SimCLR, Boyl) use a ResNet Encoder that is trained supervised in their architecture (for evaluation)?
In the Simple Siam Paper, the authors call it in their abstract: "unsupervised visual representation learning", but they (and the other mentioned methods) are using a supervised (ResNet) encoders as "backbone" which seems to me not really/fully UNsupervised. I am wrong at this point?
I am stuck at the next point: the final reported results are achieved by using a linear supervised classifier on top. Why is this not done unsupervised? For example, using kNN?
So why are this kind of work is considered as unsupervised when the only unsupervised learned parameter is a head MLP and projector MLP. Are there any results doing this fully unsupervised?
I believe that if this were really done completely unsupervised, the results would be a lot worse. So I think it's better to call it semi-supervised. That way it is easier for people who are not familiar with it to understand what is being done here. And for people who do this completely unsupervised it is also fairer, when their results are compared with these here.