Label

In supervised learning a labelled example is a pair (x, y) of a feature vector and a label; Google's ML glossary defines the label as the answer or result portion of an example. For classification y is a category from a fixed label set (binary, multi-class, or multi-label when several categories may apply at once); for regression it is a real number. Losses compare predictions to labels: cross-entropy against one-hot or class-index labels, squared error against numeric targets. Techniques such as label smoothing replace hard one-hot targets with slightly softened ones to curb overconfidence.

Labels come from human annotation, from records that already exist (a later diagnosis, whether a loan defaulted, whether a user clicked), or from heuristics and weaker models (weak supervision). Ground truth is the term for the true value against which predictions are judged; in practice the recorded label is only an estimate of it. Label noise is common: Northcutt, Athalye and Mueller (2021) estimated an average of at least 3.3 percent label errors across the test sets of ten widely used benchmarks, and at least 6 percent in the ImageNet validation set, enough to change which of two models appears better.

Labels encode decisions and can encode bias. If historical hiring outcomes or arrest records are used as labels, a model learns to reproduce those past decisions rather than the quality they were meant to measure. Proxy labels (clicks for relevance, cost of care for health need) are a frequent source of this kind of silent target mismatch.

Self-supervised learning avoids manual labels by deriving the target from the input itself, for example the next token in a text, which is why large language models can pretrain on unlabelled corpora. Later stages such as instruction tuning and RLHF bring human-provided targets back in the form of demonstrations and preference rankings.