A Simple Framework for Contrastive Learning of Visual Representations (SimCLR)¶
Authors: Ting Chen, Simon Kornblith, Mohammad Norouzi, Geoffrey Hinton Year: 2020 Venue: ICML 2020 Citations: 15,000+
Link¶
Summary¶
Introduces SimCLR, a simple yet effective contrastive learning framework for self-supervised learning. Uses data augmentation, projections, and contrastive loss to learn visual representations. Without labels, achieves competitive performance with supervised methods.
Key Concepts¶
- Self-Supervised Learning: Learn representations without labels
- Contrastive Loss: Maximize similarity between augmented views of same image
- Data Augmentation: Critical component of contrastive methods
- Projection Head: Non-linear projection for contrastive loss
- Momentum Contrast: Maintains queue of negative samples
- Pre-training: Unsupervised pre-training enabling efficient fine-tuning
Impact¶
- 15,000+ citations
- Sparked self-supervised learning revolution in vision
- Inspired MoCo, BYOL, SwAV, and many variants
- Practical alternative to supervised pre-training
- Reduced labeled data requirements for training
- Influenced NLP self-supervised methods (ELECTRA, etc.)
- Critical technique in modern vision pipelines
Key Results¶
- Self-supervised learning competitive with ImageNet supervised pre-training
- Scaling benefits with larger models
- Strong transfer learning performance
- Better generalization than supervised models
Related Papers¶
- Momentum Contrast (MoCo, He et al., 2020)
- Bootstrap Your Own Latent (BYOL, Grill et al., 2020)
- Contrastive Divergence (Hinton, 2002)
- Metric Learning (Siamese networks, Koch et al., 2015)