---
title: Supervised Learning
description: Page on Vedang Vatsa's site: https://veda.ng/glossary/supervised-learning
canonical: https://veda.ng/glossary/supervised-learning
last_updated: 2026-10-03
type: text/markdown
---
# Supervised Learning

Source: https://veda.ng/glossary/supervised-learning
Author: Vedang Vatsa (https://veda.ng/about)

Supervised learning trains a model on labeled examples where every input has a known correct output.

You show thousands of images labeled "car" and thousands labeled "bicycle." The algorithm looks at pixels, finds patterns such as wheels, windshields, or handlebars that correlate with each label, and stores that mapping. After training it classifies new unlabeled images. Email spam filters work the same way: millions of messages humans marked "spam" or "not spam." Medical diagnostics use the same method: models trained on thousands of labeled X-rays can flag early tumors or pneumonia.

The bottleneck is labeled data. High-quality human annotation is expensive and slow. That cost is why semi-supervised learning and self-supervised learning exist: they try to get similar skill from fewer labels.

The loss compares the model’s output to the label. Classification uses labels such as car versus bicycle or spam versus not spam. Medical models use radiologist marks on X-rays for tumors or pneumonia. After training, inference runs on unlabeled inputs. Accuracy is bounded by label quality and coverage. If a class is missing from the training set, the model cannot name it. That is the expensive part: people have to annotate thousands to millions of examples. Semi-supervised and self-supervised methods exist to spend fewer of those labels. Supervised learning is training on labeled pairs. Spam filters and ImageNet classifiers are the textbook examples.

Glossary index: https://veda.ng/glossary