Learning without labels.
WHAT UNSUPERVISED LEARNING DOES
Finds structure in data with no known answers.
WHAT THE COMMON TASKS ARE
- Clustering: grouping similar items
- Dimensionality reduction: representing data with fewer values
- Anomaly detection: identifying the unusual
- Association: finding items that occur together
WHAT CLUSTERING SUITS
Segmentation, where the segments are not known in advance.
WHAT MAKES IT DIFFICULT
There is no correct answer to compare against.
WHAT THAT MEANS
Evaluation is partly judgement, and cluster counts are chosen rather than discovered.
WHAT DIMENSIONALITY REDUCTION PROVIDES
Visualisation Removal of redundancy Faster training
WHAT SELF-SUPERVISED LEARNING IS
Creating labels from the data itself, such as predicting a hidden part from the rest.
WHY IT MATTERS ENORMOUSLY
It allows learning from vast unlabelled data, which is what made modern language and vision models possible.
WHAT THE PATTERN IS
Pre-train on unlabelled data, then adapt with a smaller labelled set.
WHY THAT MATTERS PRACTICALLY
Useful models become achievable with hundreds of labels rather than millions.
WHAT TO CONSIDER FIRST
Whether a pre-trained model already does most of the work.
