In plain words
A label is the answer we want a computer to learn. A photo plus the word "cat". An email plus "spam" or "not spam". A flat's details plus the price it sold for. Put the example and its answer together and you get one piece of labelled data.
Computers learning by supervised learning need lots of these. They study the examples, guess, and check each guess against the label. Without labels, there is nothing to check against.
Some labels come for free, like past sale prices in records. Many are added by people, one example at a time: drawing boxes around cars, tagging photos, rating a chatbot's answers. That work is slow and costly, and mistakes in the labels become mistakes in the model.
Everyday example
Writing names on the back of old photos
A grandmother goes through the family album and writes on the back of each photo: "Mamaji, 1984", "Priya's mundan". Years later, a young cousin studies those labelled photos and starts recognising relatives in photos nobody has marked. The photos are the data, her notes are the labels, and the cousin is the model.
Where the example breaks: The cousin can ask questions and remember stories. A model only sees the pairs it is given, so if the labels are wrong or one-sided, it learns those errors faithfully.
How it actually works
In machine learning terms, a labelled example is a set of features (the input) plus a label (the output we want predicted).
Where labels come from
Some exist already: prices in sales records, which emails users marked as spam, which ads were clicked. Others are made on purpose by human annotators following written guidelines. Companies run this as a whole workflow, often called data annotation, with people kept in the loop.
Labels at scale
ImageNet is the famous case. Its team used Amazon Mechanical Turk, and its creator Fei-Fei Li has said almost 50,000 workers in 167 countries helped clean and label the images, more than 14 million in the end. A 1.2 million image subset trained AlexNet in 2012.
Labels in chatbots
Language models use labels too. People write example answers for fine-tuning and rank pairs of answers for RLHF. Those rankings are labelled data.
Quality problems
Labellers disagree, get tired and carry their own assumptions, which can feed bias into a model. A subtler problem is label leakage: a feature that secretly gives away the answer, so the model looks brilliant in training and fails in use.
Ways to need fewer labels
Self-supervised learning makes labels from the data itself. Active learning asks people to label only the examples the model is most unsure about. Synthetic data can fill gaps.
- Also called
- Annotated data, ground truth; the work is called data labelling or annotation
- Famous dataset
- ImageNet: more than 14 million hand-annotated images in over 20,000 categories, first shown at CVPR 2009
- Needed for
- Supervised learning, test sets and evaluation, chatbot fine-tuning
- Main cost
- Human time and expertise, especially for medical or legal labels
Questions people ask
What is the difference between labelled and unlabelled data?
Labelled data has the answer attached to each example, such as a photo tagged "dog". Unlabelled data is just the raw examples. Supervised learning needs labels; unsupervised and self-supervised learning can work without them.
Who labels data for AI?
Sometimes the labels already exist in records. Otherwise people add them: in-house teams, specialist annotation companies and crowd workers around the world. For expert tasks like medical scans, doctors or trained specialists do the labelling.
Is it labelled or labeled?
Both are correct. "Labelled" is British and Indian spelling, "labeled" is American. Google's machine learning glossary, for example, uses "labeled example". They mean exactly the same thing.
Why does label quality matter so much?
Because the model treats every label as the truth. If labellers make mistakes or share a blind spot, the model learns them as patterns, and its scores on a test set labelled the same way will not reveal the problem.