Unsupervised Learning
Finding structure such as clusters or low-dimensional representations in data that has no labels.
Definition
Unsupervised learning operates on unlabeled data, discovering structure without a target output. Typical goals are grouping similar points (clustering), compressing to fewer dimensions (dimensionality reduction), and estimating the data's probability distribution (density estimation).
Self-supervised learning, a fast-growing variant, invents a supervised task from unlabeled data, such as predicting a masked word or a missing image patch. This has become the dominant way to pretrain large models, blurring the old line between supervised and unsupervised methods.
Evaluation is the perennial difficulty, since there is no label to compare against. Practitioners rely on internal measures of cluster quality, on stability across resampling, and ultimately on whether the discovered structure improves a downstream task or matches expert judgment. Because different methods encode different notions of similarity, the structure they report is best treated as a lens on the data rather than a definitive partition of it.
Because there is no ground-truth label, evaluation is indirect: cluster compactness, reconstruction error, or downstream usefulness of the learned representation.
Representative methods
- k-means and hierarchical clustering
- Principal component analysis
- Autoencoders for learned compression
- Anomaly detection as an unsupervised task
Why it matters
Labels are expensive; raw data is abundant. Unsupervised methods extract usable structure from what you already have and often serve as a pre-processing step that improves later supervised models.
Fusion connection
Clustering large ensembles of simulated Hyperion operating points reveals families of similar plasma behavior, helping engineers reason about the design space before committing to detailed study.