About other self-supervised framework
EBGU opened this issue · comments
Hi!
I read your paper and it is really great! I am also working with contrast learning applied to videos, and I am trying frameworks that don't need negative samples(like simsiam). I wonder if you have tried those frameworks? My networks suffered from either collapse or no convergence, so I just want to discuss with you.
Thank you!
Simsiam is indeed easier to collapse, you can refer to the methods of preventing collapse in the article "Exploring Simple Siamese Representation Learning".