Self-supervision manufactures the label from the input.

Demonstration

Labels are the expensive part. A radiologist’s time is the binding constraint on every medical vision dataset, and the images themselves are comparatively abundant. Self-supervision is the observation that a supervision signal can be constructed rather than collected.

The recipe: apply a transformation whose parameters you know, and ask the model to recover them.

• Hide a region, predict what was there. The answer is the region you hid.
• Take two crops of one photograph, ask whether they came from the same image. The answer is your bookkeeping.
• Rotate by a multiple of 90°, predict the angle. The answer is the angle you chose.

None of these is the task anyone cares about. They are pretext tasks, and the whole bet is that solving them is impossible without learning something genuinely useful about images. To fill in a masked region you must have some model of what usually goes there; to match two crops you must represent content rather than framing.

The bet holds unevenly, and the choice of pretext is where the design lives. Rotation prediction works on photographs of objects because photographs have a canonical orientation, and fails on a fundus image or a satellite tile because they do not — there, the task is either impossible or solvable from an artefact.

Chapter V is about the two families that turned out to scale: contrastive methods, which build the task from pairs, and masked modelling, which builds it from reconstruction.

Corollary

Judge a pretext task by asking what shortcut would solve it. Early context-prediction work was partly solved by chromatic aberration — the model learned the lens, not the scene. If a corruption leaves a cheaper cue than the one you intended, that cue is what gets learned.