Momentum Contrast for Unsupervised Visual Representation Learning (MoCo)¶
Authors: Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, Ross Girshick Year: 2020 Venue: CVPR 2020 Citations: 10,000+
Link¶
Summary¶
Introduces Momentum Contrast (MoCo), building large queues of negative samples for contrastive learning. Uses momentum encoder to update queue, enabling large-scale self-supervised visual representation learning.
Key Concepts¶
- Contrastive Learning: Maximize agreement between augmented views
- Momentum Encoder: Slowly updated copy of query encoder
- Queue of Negatives: Maintains large dictionary of negative samples
- Consistency: Momentum helps maintain consistency across batches
- Efficiency: Works with memory-efficient implementation
- Scalability: Larger negative sample sets improve learning
Impact¶
- 10,000+ citations
- Competing approach to SimCLR for self-supervised learning
- Better scalability with momentum encoder
- Influenced many subsequent contrastive methods
- Foundation for vision pre-training without labels
- Important for understanding contrastive mechanisms
Key Results¶
- Competitive with SimCLR and supervised ImageNet pre-training
- Better memory efficiency than SimCLR
- Strong transfer learning performance
- Generalizes to downstream tasks
Related Papers¶
- SimCLR: Contrastive Learning (Chen et al., 2020)
- Bootstrap Your Own Latent (BYOL, Grill et al., 2020)
- Contrastive Predictive Coding (van den Oord et al., 2018)
- Metric Learning Fundamentals (Siamese networks, 2015)