Unsupervised Learning

Unsupervised learning starts with inputs but no target labels. The model searches for structure that is already latent in the data: groups, low-dimensional geometry, recurring co-occurrences, or probability distributions. There is no answer key, so a useful result depends on the question being asked.

Clustering assigns similar examples to groups using methods such as k-means, DBSCAN, or hierarchical clustering. Dimensionality reduction maps high-dimensional observations into a smaller coordinate system while preserving important variation; PCA is useful for linear structure, while t-SNE and UMAP are mainly exploratory visualization tools. Density estimation models where data is likely to occur, which supports anomaly detection and generative sampling.

Unsupervised methods are valuable when labels are expensive or incomplete. They can reveal customer segments, organize document collections, surface unusual transactions, or provide compact features for a downstream model. The output still needs interpretation: a cluster is a statistical grouping, not automatically a real-world category.

Self-supervised learning is closely related but not identical. It creates a training target from the input itself, such as a masked token or the next token, and uses that proxy task to learn representations. This distinction matters because modern language and vision models are usually pretrained with self-supervision rather than with purely label-free clustering.