CVPR-2020 Momentum Contrast for Unsupervised Visual Representation Learning

2020-03-23 · CVPR· Vision · vision-language, contrastive-learning

Paper: Momentum Contrast for Unsupervised Visual Representation Learning

MoCo v1: Contrastive learning as dictionary lookup, with a queue and momentum encoder for a large dictionary

Strong writing craft: The first paragraph of the intro contrasts CV and NLP and explains why unsupervised learning has underperformed in CV; the second introduces contrastive learning without heavy detail, summarizes prior methods as dictionary lookup, and proposes MoCo under a unified CV/NLP framing—broadening the paper’s overall scope.
Two design constraints are validated: the dictionary should be large, and representations should be consistent.

A large dictionary forces the model to learn more discriminative features;

Consistency keeps learning in a shared, semantically aligned encoder space and helps avoid shortcut solutions.
The encoder is updated by gradient backpropagation; the momentum encoder is updated with a large momentum coefficient. A queue structure links batch size to dictionary size so the dictionary can be made sufficiently large.