self-supervised learning

Self-supervised learning is a way to train a model on images that nobody has labelled. The trick is to hide part of the data or transform it, and then make the model predict the hidden or original part. Because the answer is already inside the data, the data supervises itself, so no human annotator is needed. Think of covering half of a photograph and asking a student to imagine what was behind your hand: to succeed they must understand objects, textures and context, and that understanding is exactly the reusable knowledge we are after.

More precisely, SSL defines a pretext task whose target is derived automatically from each unlabelled input, trains a network (typically a CNN like ResNet or a Vision Transformer) to solve it, and then throws the pretext-specific output away, keeping the internal feature extractor. The features are transferred to real downstream tasks (classification, detection, segmentation) by fine-tuning or by training a small head on top. This two-stage recipe, large-scale unlabelled pretraining followed by small-scale labelled adaptation, is why SSL matters: labels are expensive and scarce, but raw images are essentially free and effectively unlimited.

Modern SSL splits into two broad families. Joint-embedding methods (contrastive ones like SimCLR and MoCo, and non-contrastive ones like BYOL, SimSiam, SwAV and DINO) push the embeddings of two augmented views of the same image to agree while avoiding a constant useless solution. Generative or denoising methods (masked image modeling such as MAE and BEiT) mask out patches and reconstruct them. Both turn the structure already present in images, spatial layout, colour, co-occurrence, into a free training signal.