Self-supervised learning is a training strategy where a model generates its own labels from raw, unlabeled data. Instead of relying on human annotators, the algorithm creates proxy tasks from the data itself. A common example is masking part of an image and asking the network to reconstruct the missing region, which forces it to learn useful visual features without any manual labels.

This approach has become central to modern computer vision because labeled datasets are expensive and slow to build. Models pre-trained with self-supervised objectives on large image collections develop strong general-purpose representations. Those representations then transfer well to downstream tasks like classification, detection, and segmentation, often matching or beating fully supervised baselines trained on far more labeled examples.

Frameworks like DINO, MAE, and SimCLR each take a different angle on self-supervision. DINO uses a teacher-student setup with augmented views, MAE masks and reconstructs image patches, and SimCLR pulls augmented pairs closer in feature space while pushing unrelated images apart. In practice, teams pre-train on unlabeled data with one of these methods, then fine-tune on a small labeled set for their specific task.