EMP-SSL

[surajpaib]

Hi Team,

I recently came across this paper (Link: EMP-SSL Paper). The authors have provided their implementation on Github (EMP-SSL Repository).

TLDR of the method: The key towards efficient self-supervised learning is to increase the number of crops from each image instance.

Key contributions:

• EMP-SSL can reduce the training epochs by two orders of magnitude.
• The method seems to transfer well to out-of-domain tasks.
• Due to the reduction in training epochs, the overall training time is considerably decreased. This, however, may come at the cost of increased memory usage.

Considering the promising results this method has shown, I think it could be a valuable addition to the collection of self-supervised learning methods so that it is accessible to the larger community.

[philippmwirth]

Hi @surajpaib thanks for bringing this up! I'll add it to the tracker #1172.

[johnsutor]

I can implement this.

[philippmwirth]

@johnsutor let me know if you need any assistance 👍

[deepsthewarrior]

I can implement this.

Are you already working on this?

[johnsutor]

@deepsthewarrior yep!