Finding Structure Without Labels | AI Learning Guide

When there are no labels, the model cannot be told it was wrong. So the feedback has to come from the data itself. The comparison on this page shows the two most common ways that happens. In grouping, the model is rewarded for putting similar items together and keeping dissimilar ones apart, and the signal is how well those groups reflect real similarity rather than arbitrary splits. In compression, the model must rebuild the data from a shortened description, and the signal is how much detail is lost. The trade-off is that neither goal has a single correct answer, so judging whether the structure is actually useful usually needs a human or a later task.

0:00 / 0:00

Auto-play

Unsupervised learning has no labels, so the feedback signal comes from the structure of the data itself — how well the model groups similar items or compresses the data without losing what matters.

Two common unsupervised goals

Grouping

Place similar items in the same group
Signal: how well groups reflect real similarity
Useful when categories are unknown

Compression

Reconstruct data from a shorter description
Signal: how much detail is lost
Useful for summarizing or denoising

Without labels there is no single correct answer to check against, so judging whether the discovered structure is meaningful usually requires a human or a downstream task.