Self-supervised learning is a training approach where models learn from unlabeled data by generating their own supervision signals. Instead of relying on manual labels, SSL defines pretext tasks—learning objectives that force the model to understand the structure or semantics of the data. Once trained, the learned representations can be transferred to downstream tasks like classification or detection.

Examples of pretext tasks include predicting masked tokens in a sentence (e.g., BERT), solving jigsaw puzzles from image patches, or matching image and text pairs (e.g., CLIP). In vision, contrastive methods like SimCLR or BYOL train models to bring different views of the same image closer in latent space while pushing apart others.

SSL has become a dominant paradigm in both NLP and computer vision, significantly reducing the need for labeled datasets while improving generalization. It's especially valuable in domains with limited annotations or high labeling costs.

By learning from the data itself, SSL shifts the focus from annotation-heavy pipelines to scalable, unlabeled learning.