Unsupervised Learning

Unsupervised learning is the branch of machine learning that looks for structure in data without labels. The model is never told the right answer; it must discover regularities, groups and patterns on its own. This sets it apart from supervised learning, which learns from already-labelled examples (this image is a cat), and from reinforcement learning, which learns from rewards earned by acting in an environment. Where supervised learning imitates and reinforcement learning experiments, unsupervised learning explores.

In the language of the scikit-learn documentation and the classic textbooks (Bishop; Goodfellow, Bengio and Courville), its raw material is the unannotated example and its goal is to model how the data is distributed or organised.

The tasks

Four families cover almost everything. Clustering gathers similar examples into groups: k-means, DBSCAN, hierarchical methods. Dimensionality reduction compresses many variables into a few while keeping the essentials, through techniques such as PCA or t-SNE, useful both for visualisation and for feeding downstream models. Density estimation models the probability of the data and underpins anomaly detection: the rare is simply the improbable. And representation learning seeks compact, rich encodings —autoencoders— that capture latent meaning. IA360 keeps separate entries for clustering and dimensionality reduction; here it is enough to see them as pieces of one map.

Self-supervision and foundation models

The defining shift of the past decade has a name: self-supervised learning. In his 2016 NeurIPS invited talk, Yann LeCun presented learning from raw data as the major unsolved challenge; his familiar “cake analogy” made unsupervised learning the cake, supervised learning the icing, and reinforcement learning the cherry. LeCun and Ishan Misra later explained why self-supervised is the more precise term: the system obtains supervisory signals from the structure of the data itself. The idea is to derive the objective from the data. BERT hides tokens and trains the model to reconstruct them; the first GPT maximizes the probability of the next token given its preceding context. Unannotated text thus yields many exercises whose answers are already present in the text itself. That pretraining is now the engine of foundation models: the Stanford report that coined the term (Bommasani et al., 2021) defines them as models trained with self-supervision at scale on broad data and then adapted to a wide range of downstream tasks. Documentation: representation-learning chapter; original BERT paper; original foundation-model report.

Uses and limits

It serves to segment customers, flag fraud and anomalies, compress information and learn reusable representations. Its honest limit is that evaluation without labels is difficult. The clustering-evaluation documentation distinguishes metrics that require reference classes from internal measures that inspect only the data and the partition. A measure such as the silhouette quantifies cohesion and separation, but it cannot decide whether the groups answer the useful question. Clustering quality therefore calls for several metrics plus domain validation, while a representation is often tested through what it enables downstream. Comparing two solutions without ground truth requires an explicit criterion; no single score certifies meaning and utility at once.