What is SimCLR and how does contrastive learning work in vision?

SimCLR is a self-supervised learning framework that learns representations by contrasting positive and negative image pairs.

SimCLR applies strong data augmentations to create two views of the same image (positive pair) and treats other images as negatives. A CNN encoder and projection head map images into embedding space where contrastive loss (NT-Xent) maximizes similarity between positives and minimizes similarity between negatives.