Supervised, Unsupervised, and Reinforcement Learning

Not all machine learning works the same way. The three main approaches differ in what kind of feedback the model receives during training.

Supervised learning

The model is trained on labelled examples — inputs paired with correct answers. "This photo is a cat. This photo is a dog. Now predict what new photos show." Most practical AI applications use supervised learning: spam detection, image recognition, medical diagnosis support. The challenge is that labelling data is expensive and time-consuming.

Unsupervised learning

The model finds patterns in unlabelled data on its own, without being told what to look for. This is useful for discovering hidden structure — clustering customers by behaviour, finding topics in large document collections, detecting unusual transactions (fraud detection). The model groups or organises things, but a human must interpret what those groupings mean.

Reinforcement learning

The model learns by trial and error in an environment, receiving rewards for good actions and penalties for bad ones. This is how AI has learned to play games — from Chess to Go to StarCraft — eventually surpassing the best human players. It is also how many LLMs are fine-tuned to be more helpful and safe after initial training.

Which one to use?

The choice depends on what data is available and what the task is. In practice, many systems combine approaches: a model might be pre-trained with supervised learning on labelled data, then fine-tuned with reinforcement learning based on human feedback.

Understanding which type of learning a system uses often explains its limitations. Supervised learning needs labelled data. Reinforcement learning needs a clear reward signal. Knowing this helps you ask the right questions about any AI system.