The leading representation learning methods, SimCLR and Moco, should utilize a large number of negative samples and so both require huge batch sizes. However, PCL, which is based on EM algorithm uses cluster centers of samples instead of large number of negative ones , solves this problem, and surpasses Moco and others.
TL;DR
The leading representation learning methods, SimCLR and Moco, should utilize a large number of negative samples and so both require huge batch sizes. However, PCL, which is based on EM algorithm uses cluster centers of samples instead of large number of negative ones , solves this problem, and surpasses Moco and others.
Why it matters:
Paper URL
https://arxiv.org/abs/2005.04966
Submission Dates(yyyy/mm/dd)
Authors and institutions
Methods
Results
Comments