What Is Machine Learning: Supervised, Unsupervised, and Semi-Supervised

Learning Without Explicit Programming

Traditional software follows instructions written by human programmers: if this, then that. Machine learning takes a fundamentally different approach. Instead of encoding rules, a machine learning system is given data — examples of inputs and outputs, or just inputs alone — and it discovers the rules itself by optimizing a mathematical objective. The programmer provides the architecture, the training data, and the objective; the algorithm finds the parameters that best satisfy the objective on that data. The result is a model that can generalize from the examples it has seen to new examples it has never encountered.

This paradigm shift opened up categories of problems that explicit programming could never address. There is no way to write rules for recognizing every possible face, transcribing every accent, or predicting which loans will default — the variation is too high and the relevant features too subtle. But machine learning systems can solve all of these problems if trained on sufficient data. Understanding the three main learning paradigms — supervised, unsupervised, and semi-supervised — is the starting point for understanding how machine learning achieves this.

Supervised Learning: Learning from Labels

Supervised learning is the most widely deployed paradigm. The training data consists of input-output pairs: each example includes both the input (an image, a sentence, a row of tabular data) and the corresponding label (the object category, the sentiment, the purchase decision). The model learns a function that maps inputs to outputs by minimizing the difference between its predictions and the provided labels. The word "supervised" refers to the fact that the correct answers are provided, supervising the learning process.

Classification and regression are the two main flavors. In classification, the output is a discrete category — spam or not spam, cat or dog, fraudulent or legitimate. In regression, the output is a continuous number — a stock price, a house value, a temperature forecast. The same algorithms (logistic regression, decision trees, support vector machines, neural networks) can often be adapted for both tasks by changing the output layer and loss function. Applications of supervised learning are ubiquitous: spam filters, medical diagnosis systems, credit scoring models, image classifiers, and speech recognition are all trained with labeled data.

The Supervised Learning Pipeline

Building a supervised learning system involves several steps beyond choosing an algorithm. Data collection and labeling is often the most expensive and time-consuming part. Human annotators may need to label millions of images, transcribe thousands of hours of speech, or rate the sentiment of countless reviews. Data labeling platforms and crowdsourcing marketplaces have grown into substantial industries to meet this demand. The quality and representativeness of labels directly determines what the model can learn — biased or noisy labels produce biased or noisy models.

After labeling, data is split into training, validation, and test sets. The model trains on the training set, hyperparameters are tuned using the validation set, and final performance is reported on the test set (which the model never sees during training or tuning). This split prevents the model from being evaluated on data it has memorized. Feature engineering — selecting, transforming, and combining raw input variables — was historically crucial but has been largely automated by deep learning models that learn representations directly from raw data.

Unsupervised Learning: Discovering Structure Without Labels

Unsupervised learning operates on data that has no labels. The goal is not to predict a target variable but to discover inherent structure in the data itself — clusters, dimensions, distributions, or generative factors. Because labeling is expensive and time-consuming, unlabeled data is far more abundant; unsupervised methods unlock the value of this vast resource.

Clustering algorithms like k-means and DBSCAN group data points that are similar to one another, without being told what the groups mean. These are used for customer segmentation, document organization, anomaly detection, and biological taxonomy. Dimensionality reduction methods like principal component analysis (PCA) and t-SNE compress high-dimensional data into two or three dimensions while preserving as much structure as possible, enabling visualization and removing redundant features. Generative models — including variational autoencoders and generative adversarial networks — learn the underlying probability distribution of the training data well enough to sample new, realistic examples from it. Unsupervised pre-training, where a model is trained on large unlabeled corpora before being fine-tuned on labeled data, is central to the modern large language model paradigm.

Reinforcement Learning: Learning from Interaction

A third major paradigm, often contrasted with supervised and unsupervised learning, is reinforcement learning (RL). An RL agent learns by interacting with an environment, taking actions and receiving scalar reward signals that indicate how well it is doing. The agent's goal is to discover a policy — a mapping from observed states to actions — that maximizes cumulative reward over time. Unlike supervised learning, there is no labeled dataset; the agent must explore the action space to discover which actions lead to high reward, balancing exploration of uncertain options against exploitation of known good ones.

Reinforcement learning has produced some of AI's most dramatic demonstrations: DeepMind's AlphaGo and AlphaZero mastered Go and chess at superhuman levels through self-play RL; OpenAI Five defeated professional Dota 2 teams; robotic systems learn to grasp novel objects through millions of simulated trials. RL is also central to RLHF (reinforcement learning from human feedback), the technique used to align large language models with human preferences — making them more helpful, less harmful, and more honest than the raw pre-trained models.

Semi-Supervised and Self-Supervised Learning

Semi-supervised learning occupies the space between supervised and unsupervised methods. It uses a small amount of labeled data alongside a large amount of unlabeled data. The labeled examples provide the target signal, while the unlabeled examples help the model learn better representations of the input space. Consistency regularization methods, for instance, encourage the model to produce the same output for an unlabeled example and a slightly perturbed version of it, exploiting the smoothness of the data manifold. This is especially valuable in domains like medical imaging where labels are expensive to obtain but unlabeled images are plentiful.

Self-supervised learning is a variant of unsupervised learning in which labels are derived automatically from the data itself rather than provided by humans. The model is trained to predict some part of the input from the rest: predict the next word in a sentence (the GPT objective), predict a masked word from its context (the BERT objective), predict one view of an image from another (contrastive learning methods like SimCLR and DINO). These pretext tasks force the model to develop rich internal representations, which can then be fine-tuned efficiently on small labeled datasets. Self-supervised learning is the foundation of the pre-train-then-fine-tune paradigm that dominates modern AI.

Choosing the Right Paradigm

Practitioners choose among these paradigms based on what data is available and what the goal is. Supervised learning is the default when high-quality labeled data can be assembled and a specific prediction task is defined. Unsupervised methods are used for exploration, compression, anomaly detection, and leveraging large unlabeled corpora. Reinforcement learning is applied in sequential decision-making problems with a clear reward signal. Semi-supervised and self-supervised methods offer a middle path when labeled data is scarce but unlabeled data is abundant.

In practice, modern AI systems often combine multiple paradigms. A language model is pre-trained with self-supervised next-token prediction on unlabeled text, then fine-tuned with supervised learning on instruction-following examples, then refined with reinforcement learning from human feedback. A recommendation system may use unsupervised clustering to segment users, supervised learning to predict click-through rates, and reinforcement learning to optimize long-term engagement. Understanding each paradigm's assumptions, strengths, and failure modes is essential for building systems that work reliably in the real world.