Momentum Contrast for Unsupervised Visual Representation Learning
Abstract
Sowe Ebou A* and Bah Yunusa A
This brief report presents a novel unsupervised learning representation learning method called momentum contrast. Momentum contrast uses a contrastive learning technique to learn representations by comparing features of related yet dissimilar images for efficient feature extraction and unsupervised representation learning. Similar images are grouped together, and dissimilar images are placed far apart. The method builds upon previous works in contrastive learning but includes a momentum optimisation step to improve representation learning performance and generate better quality representations. Experiments on various datasets demonstrate that momentum contrast is able to learn high-quality representations, allowing us to directly use them to achieve competitive performance with fewer labelled examples.