Supervised vs Unsupervised vs Reinforcement Learning

Supervised, unsupervised, and reinforcement learning are the three broad paradigms machine learning falls into, distinguished by what kind of feedback the model gets while it learns. Supervised learning trains on labeled examples with known correct answers. Unsupervised learning finds structure in data that has no labels at all. Reinforcement learning learns from a reward signal produced by acting in an environment, with no example of the “right” answer ever provided directly. Most systems you interact with combine more than one of these, but the distinction between them is foundational to understanding how any given model was actually trained.

Supervised learning: learning from labeled examples

In supervised learning, every training example comes with a known correct output — a label. An email is marked spam or not spam; an image is tagged with the object it contains; a house’s features are paired with its actual sale price. The model’s job during training is to adjust its internal parameters so its predictions get closer to those known labels, typically through gradient descent minimizing a loss function that measures how wrong the current predictions are, with backpropagation computing how to adjust each parameter to reduce that error.

Supervised learning covers most of what people mean by “traditional” machine learning: classification (which category does this belong to) and regression (what number does this map to). Its main constraint is that it needs labeled data, and labeling is often the most expensive part of the whole pipeline — someone or something has to produce the correct answer for every training example before the model can learn from it.

Unsupervised learning: finding structure without labels

Unsupervised learning gets no labels at all — just raw data — and the task is to find structure in it. Clustering algorithms group similar data points together without being told what the groups should represent. Dimensionality reduction techniques compress high-dimensional data down to a smaller number of informative dimensions. Anomaly detection flags data points that don’t fit the patterns the rest of the data shows.

Because there’s no ground truth to check against, evaluating an unsupervised model is inherently fuzzier than evaluating a supervised one — there’s no single “correct” clustering of a dataset the way there’s a correct label for a spam classifier. What makes a clustering useful is usually judged by whether it’s actionable for the task at hand, not by comparing against a known answer. Generative models that learn to produce new data resembling their training set — including GANs — sit largely in this category, learning the underlying distribution of the data rather than mapping inputs to fixed labels.

Reinforcement learning: learning from reward, not labels

Reinforcement learning is structured differently from both of the above. There’s no dataset of labeled examples and no fixed dataset to find structure in — instead, an agent takes actions in an environment and receives a reward signal indicating how good or bad the outcome was. The agent’s objective is to learn a policy — a strategy for choosing actions — that maximizes cumulative reward over time, discovering that policy through trial and error rather than being shown correct answers directly.

This paradigm is a natural fit for sequential decision-making problems: game playing, robotics, and — increasingly — aligning large language models with human preferences. RLHF applies reinforcement learning to LLM training, using a reward model trained on human preference judgments as the reward signal the policy is optimized against. It’s also the paradigm where reward hacking shows up as a distinct failure mode — an agent finding a way to maximize the reward signal that technically satisfies it without producing the behavior the reward was actually meant to encourage, because the reward function is a proxy for the real goal, not the goal itself.

Comparing the three

	Supervised	Unsupervised	Reinforcement
Feedback signal	Known correct labels	None — raw, unlabeled data	Reward signal from an environment
Goal	Predict labels for new inputs	Discover structure or patterns	Learn a policy maximizing cumulative reward
Typical tasks	Classification, regression	Clustering, dimensionality reduction	Sequential decision-making, control, RLHF
Data requirement	Labeled examples	Unlabeled data	An environment and a reward function
Evaluation	Compare predictions to known labels	Often qualitative or task-dependent	Total reward accumulated over time

Where these boundaries blur

Real systems rarely stick to one paradigm in isolation. A large language model’s pretraining phase — predicting the next token in a sequence — is technically a form of supervised learning, since each token has a “correct” next token drawn from the training text. Fine-tuning that same model afterward often mixes supervised fine-tuning on curated examples with an RLHF stage that’s explicitly reinforcement learning. Self-supervised learning — a term used heavily in modern AI — sits at the boundary of supervised and unsupervised: the labels are generated automatically from the data itself (predicting a masked word, or the next frame in a video) rather than assigned by a human, so it has supervised learning’s training mechanics without requiring hand-labeled data.

Overfitting and underfitting are also shared risks across all three paradigms, not unique to supervised learning — see overfitting vs underfitting for how a model can fail to generalize regardless of which paradigm trained it, and neural network pruning for one technique used across paradigms to reduce a trained model’s size without a proportional loss in performance.

The takeaway

The three paradigms differ in what kind of feedback a model learns from: supervised learning uses known correct labels, unsupervised learning finds structure with no labels at all, and reinforcement learning learns a policy from a reward signal generated through interaction with an environment. Most production AI systems, especially modern LLMs, combine several of these across different training stages rather than relying on just one — but knowing which paradigm a given stage belongs to is usually the fastest way to understand what data it needed and what failure modes to expect from it.