Supervised learning

Supervised learning is training a model on input-answer pairs, so it learns to produce the answer for inputs it has never seen.

It is learning with an answer key. A child learning fruit is shown a mango and told "mango", shown a banana and told "banana", corrected when wrong. After enough rounds, the child names fruit nobody has shown before. The "supervision" is the stream of correct answers — labels — provided by someone who already knows.

inputs with answers          training              new input
(photo, "mango") × 10,000  ───────────→  model  →  photo → "mango" (93% sure)

The two big task families: classification, where the answer is a category (spam or not), and regression, where it is a number (tomorrow's electricity demand). Spam filters, medical image screening, price prediction, and credit scoring are all supervised systems.

Its strength is focus: given good labels, it is the most reliable and best-understood way to teach a machine a specific mapping. Its bottleneck is those same labels — someone must produce millions of correct answers, which is slow and costly (see data-labelling). That bottleneck drove the alternatives: unsupervised-learning uses no answers, and self-supervised-learning manufactures answers from the data itself — which is how LLMs escape the labelling cost.

Where to go next#

• Full lesson: Supervised learning
• Related terms: label, unsupervised-learning, classification, regression