What is the core idea behind unsupervised learning?
Unsupervised learning has no answer key, it discovers structure in data on its own.
How does unsupervised learning differ from related concepts?
| Concept | Difference |
|---|---|
| Unsupervised vs Supervised Learning | Supervised learning uses labelled answers. Unsupervised learning works on unlabelled data. |
| Unsupervised vs Self-Supervised | Self-supervised learning creates labels from the data itself, and is how large language models are pre-trained. |
| Unsupervised Learning vs Clustering | Clustering is one unsupervised technique. Unsupervised learning also includes dimensionality reduction and anomaly detection. |
How does unsupervised learning work?
- The model receives data with no labels
- It looks for similarities, groupings, or structure
- It organises the data into clusters or lower-dimensional representations
- No correct answers are provided to check against
What are common unsupervised learning tasks?
- Clustering similar items together
- Dimensionality reduction for visualisation and compression
- Anomaly and fraud detection
- Learning embeddings that capture meaning
Why is unsupervised learning important?
Unsupervised learning matters because most of the world's data is unlabelled. It underpins recommendation, anomaly detection, and the representation learning that makes embeddings and modern AI possible.
How is unsupervised learning used in practice?
It is used for customer segmentation, anomaly detection, recommendation, and topic discovery. Its challenge is evaluation, because without labels there is no simple measure of whether the discovered structure is correct or useful.
Frequently Asked Questions
What is an example of unsupervised learning?
Grouping customers into segments based on behaviour, without any predefined categories, is a classic example of unsupervised clustering.
Is unsupervised learning used in large language models?
Their pre-training is self-supervised, a close relative that generates its own labels from raw text, which is why LLMs can learn from vast unlabelled data.