← AI Terminology

Supervised Learning

Supervised learning is the dominant machine learning paradigm where a model is trained on labelled examples — pairs of (input, correct output) — learning a mapping function that generalises to predict correct outputs for new, unseen inputs.

It is the approach behind most deployed AI applications.
Why It Matters in AI
Supervised learning is what most people mean when they say "machine learning": given enough labelled examples, teach a model to recognise patterns that generalise. Spam detection, fraud classification, medical diagnosis from scans, image recognition, and language translation are all supervised learning problems. Its limitation — requiring labelled data — drives the development of self-supervised and semi-supervised alternatives; but for well-defined tasks with available labels, supervised learning remains the highest-accuracy approach.
Key Points
Aspect Description
Regression Predict a continuous value — house price, temperature, stock return
Fine-tuning Foundation model (self-supervised pre-training) + supervised fine-tuning on labelled task data
Classification Predict a discrete class label — spam/not spam, cat/dog, disease A/B/C
Generalisation The core goal: perform well on unseen test data, not just memorise training examples
Training signal Loss function measures error between prediction and label → gradients → weight updates
Label bottleneck Requires human-labelled data — the primary cost constraint; labels are the scarce resource
Simple Analogy
A teacher showing a student hundreds of maths problems with worked solutions (labelled training data): the student learns the rules that map questions to answers. When given a new, unseen problem (test data), they apply those learned rules — that is supervised learning. The quality of teaching (label quality) determines how well the student generalises.
Common Usage Examples
  • RandomForestClassifier().fit(X_train, y_train) — tabular classification
  • model(images).loss.backward() — neural network trained on labelled image-class pairs
  • Fine-tuning: AutoModelForSequenceClassification.from_pretrained("bert-base-uncased") on sentiment labels
  • keras.losses.SparseCategoricalCrossentropy() — classification loss for supervised multi-class training
  • ImageNet: 1.2M images with 1,000 class labels — the canonical supervised image classification benchmark
Summary
In short: Supervised learning trains models on labelled input-output pairs — the dominant ML paradigm powering most deployed AI applications, limited by the cost of human annotation and the availability of high-quality labelled datasets.