This paper introduces SimCLR, a simple framework for contrastive learning of visual representations. SimCLR simplifies recent contrastive self-supervised learning algorithms by not requiring specialized architectures or a memory bank. The framework systematically studies key components, showing that data augmentation composition is crucial for effective predictive tasks, a learnable nonlinear transformation between representations and contrastive loss improves representation quality, and contrastive learning benefits from larger batch sizes and more training steps compared to supervised learning. SimCLR outperforms previous methods on ImageNet, achieving 76.5% top-1 accuracy with a linear classifier, a 7% improvement over previous state-of-the-art. When fine-tuned on 1% of labels, it achieves 85.8% top-5 accuracy, outperforming AlexNet with 100× fewer labels. The framework uses random crop and color distortion for data augmentation, which is critical for learning generalizable features. It also employs a nonlinear projection head to improve representation quality. The contrastive loss function used is NT-Xent, which is effective for contrastive learning. The framework is evaluated on various datasets and shows strong performance in transfer learning and semi-supervised learning. The results demonstrate that SimCLR is a simple yet effective method for contrastive learning of visual representations.This paper introduces SimCLR, a simple framework for contrastive learning of visual representations. SimCLR simplifies recent contrastive self-supervised learning algorithms by not requiring specialized architectures or a memory bank. The framework systematically studies key components, showing that data augmentation composition is crucial for effective predictive tasks, a learnable nonlinear transformation between representations and contrastive loss improves representation quality, and contrastive learning benefits from larger batch sizes and more training steps compared to supervised learning. SimCLR outperforms previous methods on ImageNet, achieving 76.5% top-1 accuracy with a linear classifier, a 7% improvement over previous state-of-the-art. When fine-tuned on 1% of labels, it achieves 85.8% top-5 accuracy, outperforming AlexNet with 100× fewer labels. The framework uses random crop and color distortion for data augmentation, which is critical for learning generalizable features. It also employs a nonlinear projection head to improve representation quality. The contrastive loss function used is NT-Xent, which is effective for contrastive learning. The framework is evaluated on various datasets and shows strong performance in transfer learning and semi-supervised learning. The results demonstrate that SimCLR is a simple yet effective method for contrastive learning of visual representations.