01Three families, decided by one thing

When people say "machine learning," they usually mean one of three broad paradigms. They aren't different algorithms so much as different learning settings — and the thing that separates them is the feedback signal each one learns from. Supervised learning gets labels (the right answer attached to every example). Unsupervised learning gets no labels at all and has to find structure on its own. Reinforcement learning gets a reward — a score earned by acting in an environment over time. Keep that one question in mind — "what does it learn from?" — and the rest follows.

• The paradigms differ chiefly in their feedback signal: explicit answers (labels), no labels at all, or a scalar reward earned over time.
• The boundaries can blur — semi-supervised and self-supervised learning mix labeled and unlabeled data and don't fit cleanly into the three-way split.

02Supervised — learning from labeled answers

Supervised learning trains a model on labeled examples — input–output pairs — so it learns a mapping from inputs to known target outputs. You give it thousands of emails already marked "spam" or "not spam," and it learns the pattern that connects an email to its label, so it can label new ones it has never seen. There are two classic shapes: classification, where the answer is a discrete label (spam / not spam, cat / dog), and regression, where the answer is a continuous number (tomorrow's temperature, a house price). The whole point is to learn a mapping that generalizes — which is why supervised models are always tested on held-out data they didn't train on.

• Needs labels: every training example comes with its correct answer attached.
• Two task types: classification (discrete labels) and regression (continuous values).
• Example methods: linear and logistic regression, support vector machines, decision trees, and neural-network classifiers.

03Unsupervised — finding structure with no labels

Unsupervised learning works on unlabeled data and its job is to discover structure within it — nobody hands it the right answers. The most common version is clustering: grouping similar data points together, like sorting a pile of customer records into segments that behave alike, without anyone defining the segments in advance. The other big version is dimensionality reduction (and density estimation): compressing data into a simpler representation that keeps what matters and drops the noise. Because there's no answer key, success is harder to measure than in supervised learning — you're looking for useful patterns, not a known target.

• No labels: the system is given raw data and must find the patterns itself.
• Two big families: clustering (grouping by similarity) and dimensionality reduction / density estimation (compressing or representing data).
• Example methods: k-means and DBSCAN clustering, and PCA for dimensionality reduction.

A note for the curious: self-supervised learning — which underpins modern large language models — generates its own training signal from unlabeled data and is usually described separately from this classic unsupervised setting.

04Reinforcement — learning by reward and trial

Reinforcement learning is the odd one out. There's no fixed dataset of answers at all. Instead, an agent takes actions in an environment and learns a policy — a strategy for what to do — by maximizing a cumulative reward signal it earns through trial-and-error interaction over time. Think of a program that plays a game thousands of times, rewarded when it wins and penalized when it loses, gradually shifting its behavior toward whatever earns more reward. A defining feature is the exploration-versus-exploitation trade-off: the agent has to balance trying new actions to gather information against exploiting the actions it already knows pay off. Landmark examples include systems that learned to play Atari games directly from the screen, and the program that mastered the board game Go through self-play.

• No answer key: feedback is a scalar reward earned by acting, not a label attached to each example.
• Learns a policy: a strategy mapping situations to actions, tuned to maximize reward over time.
• Exploration vs exploitation: balancing new actions (to learn) against known good actions (to score).
• Example methods: Q-learning, deep Q-networks (DQN), and policy-gradient methods.

05Sort the task: which paradigm fits?

Here's where it clicks. Below are real machine-learning tasks. Pick a task, then choose the bin you think it belongs in — Supervised, Unsupervised, or Reinforcement. You'll get instant feedback and a one-line reason for each, and the panel underneath shows the signal each paradigm learns from. The fastest way to decide is to ask: does this task come with labeled answers, just raw data to find structure in, or a reward earned by acting?

06Check your understanding

07Take it with you & go deeper

★Sources & further reading