← AI Terminology
Unsupervised Learning
Unsupervised learning is a machine learning paradigm where models learn patterns, structure, or representations from unlabelled data — discovering hidden structure (clusters, embeddings, generative distributions) without being given correct output labels.
It is the foundation of self-supervised learning, generative AI, and representation learning.
It is the foundation of self-supervised learning, generative AI, and representation learning.
Why It Matters in AI
Labelling data is expensive and limits supervised learning to narrow domains. Unsupervised learning exploits the vastly larger pool of unlabelled data: web pages, images, molecules, and genomic sequences all exist in abundance without labels. K-means and DBSCAN cluster customers and documents; PCA and UMAP visualise high-dimensional data; autoencoders learn compact representations; generative models learn the full data distribution. Most of modern AI — pre-trained LLMs, diffusion models, embeddings — is built on unsupervised or self-supervised objectives.
Key Points
| Aspect | Description |
|---|---|
| Clustering | Group similar examples without labels — K-means, DBSCAN, hierarchical clustering |
| Contrastive | Learn embeddings by comparing similar/dissimilar pairs — SimCLR, MoCo, CLIP |
| Self-supervised | Special case of unsupervised: labels derived from data itself (masked tokens, rotation prediction) |
| Anomaly detection | Model normal distribution; flag outliers — isolation forest, one-class SVM, autoencoders |
| Generative models | Learn data distribution and generate new samples — GANs, VAEs, diffusion models, LLMs |
| Dimensionality red | Learn compact representations — PCA, t-SNE, UMAP, autoencoders |
Simple Analogy
A librarian who organises thousands of unread books without any catalogue: they group books by theme (clustering), create an index system (dimensionality reduction), and learn what a "typical book" looks like (generative model) — all without reading a category label. Patterns emerge from content alone.
Common Usage Examples
KMeans(n_clusters=5).fit(X)— cluster customer segments from purchase historyUMAP(n_components=2).fit_transform(embeddings)— visualise 1,536-dim embeddings in 2DIsolationForest().fit(X).predict(X_new)— anomaly detection on network traffic- Autoencoder:
encoder = Model(inputs, latent); decoder = Model(latent, reconstructed)— learn compression - BERT pre-training: masked language model — next sentence prediction — self-supervised (a form of unsupervised)
Summary
In short: Unsupervised learning discovers structure in data without labels — the paradigm underlying clustering, embeddings, anomaly detection, and generative AI, enabling models to learn from the vast supply of unlabelled data that supervised learning cannot access.