CVPR-2020 Improved Baselines with Momentum Contrastive Learning

· CVPR· · vision-language

Paper: Improved Baselines with Momentum Contrastive Learning

MoCo v2: Stronger baselines via a non-linear projection head and improved data augmentation

  1. Incorporates SimCLR’s MLP projection head and stronger data augmentations, together with a cosine learning-rate schedule and longer training, yielding improved results.
  2. Compared with SimCLR, MoCo v2 reaches better performance in less time while using less GPU memory and smaller batch sizes.

<HR align=left color=#987cb9 SIZE=1>