← AI Terminology

Label

A label is the ground-truth output value associated with a training example — the correct answer that a supervised learning model is trained to predict, whether a class name, a number, a bounding box, or a segmentation mask.

Without labels, supervised learning cannot occur.
Why It Matters in AI
Label quality determines model quality — "garbage in, garbage out" applies absolutely. Labelling is often the most expensive, time-consuming part of a supervised ML project: ImageNet required 14 million human-annotated images; medical AI datasets require expert radiologists. Label noise (incorrect labels), label ambiguity, and label imbalance are the leading sources of model underperformance in production. Much of modern ML (self-supervised, semi-supervised) exists specifically to reduce labelling requirements.
Key Points
Aspect Description
Regression Continuous label: a numeric value (e.g. house price, temperature) — direct float target
Label noise Mislabelled examples degrade model quality; techniques like confident learning detect them
Soft labels Probability distribution over classes (e.g. 0.8 cat, 0.2 dog) — used in distillation/label smoothing
Classification Discrete label: class name or integer (e.g. "cat", 0, 1) — often one-hot encoded
Annotation tools Label Studio, CVAT, Scale AI, Labelbox — tools for creating labelled datasets at scale
Weak supervision Programmatic labelling (Snorkel) — generate noisy labels from rules to reduce hand-labelling
Simple Analogy
Flashcards: the question is the input, the answer written on the back is the label. A student (model) studies the flashcards to learn the mapping from questions to answers. Wrong answers on the cards mean the student learns wrong patterns — label quality is everything.
Common Usage Examples
  • dataset = [{"image": img, "label": 3}, ...] — integer class label in a classification dataset
  • ImageNet: 1,000 class labels hand-annotated by Amazon Mechanical Turk workers
  • y_train = [0, 1, 1, 0, ...] — binary labels for sentiment (negative/positive)
  • COCO annotations: labels are JSON dicts with class name, bounding box, and segmentation polygon
  • Label smoothing: loss = CrossEntropyLoss(label_smoothing=0.1) — softens hard labels to improve calibration
Summary
In short: A label is the correct answer attached to each training example — the ground truth that supervises learning, and whose quality directly determines model performance.