SimCLR

Imagine learning what makes two photos depict the same thing without anyone supplying a label. SimCLR does this by showing a model two deliberately altered views of one image—such as two different crops of a dog—and teaching it that they belong together.

How SimCLR learns
SimCLR, short for Simple Framework for Contrastive Learning of Visual Representations, is a self-supervised method: it learns useful image features from the images themselves rather than from annotated classes. For every training image, it creates two random augmented views using transformations such as:

• random cropping and resizing,
• color jittering or grayscale conversion,
• Gaussian blur, and
• horizontal flipping.

Both views pass through the same neural-network encoder, such as a ResNet or Vision Transformer, then through a small projection head. The training objective, called NT-Xent contrastive loss, pulls the representations of the two views closer together while pushing them away from representations of other images in the batch.

What it teaches the model
The key idea is not to memorize pixels. A useful representation should recognize that a tightly cropped, blurred photo of a bicycle still shows the same bicycle, while distinguishing it from a car or a person. The choice of augmentations is crucial: they define which visual changes the model should ignore. After pretraining, the projection head is discarded and the encoder’s features are reused for tasks with limited labels.

Why it matters in vision
SimCLR can turn a large collection of unlabeled product photos, medical scans, or street images into a strong starting point for classification, object detection, or segmentation. This reduces dependence on expensive manual labeling and improves performance when labeled examples are scarce. Its original design benefits greatly from large batch sizes, because each batch supplies many “different-image” negatives; later methods such as MoCo address that limitation with a memory queue.