Supervised learning: the target defines the task

Supervised learning fits a mapping from inputs to known targets. A target can be a class label, a numerical value or an output sequence. The loss compares the prediction with that target and guides parameter updates.

“Supervised” describes the learning signal, not a human watching every training step. Labels can come from measurements, existing records or earlier annotation.

The same input can teach different things

A document can be paired with a topic label, a translation or a summary. Those pairings create different learning problems even though the source text is identical.

The target also limits what the score proves. A classifier trained to reproduce historical decisions may learn their systematic errors alongside useful patterns. Low loss establishes agreement with the targets, not that those targets match the real goal.

Self-supervised learning uses a related prediction setup but constructs targets from the data itself, such as hiding a word and predicting it. The distinction concerns how the supervision is obtained.