Unsupervised: finding structure

Now take away the answers entirely. In unsupervised learning there are no labels — just a pile of data — and the model's job is to find the structure hiding inside it. No one says what the groups are; the model discovers them.

Scroll to watch order emerge from an unlabeled scatter.

Two big jobs: cluster and compress

Unsupervised learning mostly does one of two things:

• Clustering — group similar items together (the animation above). "Which of my customers behave alike?" "Which of these documents are about the same topic?" The algorithm above is essentially k-means: guess centers, assign, re-center, repeat.
• Dimensionality reduction — squeeze high-dimensional data down to a few meaningful axes you can visualize or feed onward. This is exactly what the embeddings course is about: learning a compact space where similar things sit close.

The upside and the fuzziness

The huge advantage: no labeling needed, so it works on the vast oceans of raw, untagged data the world produces. The tradeoff: without answers to check against, results are harder to evaluate and easier to over-read. It's brilliant for exploring and organizing data, and it quietly powers recommendations, anomaly detection and search.

We've now seen learning from answers and learning from raw structure. Both learn from a fixed dataset that just sits there. The third paradigm is different in kind: the learner acts, and learns from the consequences. Next: reinforcement learning.