What is clustering?
Clustering is an unsupervised technique that partitions data into groups of similar items without using labels. Applied to embeddings of images or other data, it reveals the natural groupings in a dataset, which might correspond to visual themes, scenes, or hidden subpopulations you did not know were there.
It is a core tool for understanding what is actually in a large, unlabeled dataset before you commit to labeling or training.
Key takeaways
- Clustering groups similar samples without using labels.
- On embeddings, it exposes a dataset's natural structure.
- It guides exploration, sampling, and labeling decisions.
How it works
Samples are represented as vectors, usually embeddings, and a clustering algorithm groups them by proximity in that space. Methods differ in how they define groups, from centroid-based approaches like k-means to density-based ones that find arbitrarily shaped clusters and mark noise. The results are typically viewed alongside dimensionality reduction to make the structure visible.
Why it matters
You cannot curate what you do not understand, and clustering is one of the fastest ways to get a map of a large dataset. It highlights over- and under-represented groups, surfaces redundancy, and helps you sample a balanced, diverse subset to label rather than labeling at random.
Frequently asked questions
What is the difference between clustering and classification?
Classification assigns predefined labels using labeled training data, while clustering discovers groups on its own without any labels.
How is clustering used in data curation?
It reveals the structure of unlabeled data, helping you find redundant groups, rare subpopulations, and a diverse subset worth labeling.
Related terms