Supervised and Unsupervised ML
Level 3 · Kinds of learning
Where a model gets its right answers. Supervised and unsupervised learning, and the two kinds that sit in between them.
Course contentsLesson 6 of 18
Level 1 · How a machine learns
Level 2 · Getting the data ready
Level 3 · Kinds of learning
Level 4 · Classic models
Level 5 · Neural networks
Level 6 · Better models, real tools
After this lesson you can
- Tell supervised from unsupervised learning
- Say when semi-supervised learning helps
- Explain how self-supervised learning makes its labels
12 mincovers 4 glossary entries
Worth reading first: Machine Learning Basics
Where the right answers come from
Supervised learning learns from examples that come with the right answers. Unsupervised learning gets no answers at all. It looks for shape in the data by itself. Most machine learning in business is the first kind: spam filters, price guesses and loan checks. The second kind finds groups of shoppers, or odd payments that nobody tagged.
The question that sorts them is where the labels come from. In supervised learning, people supply them, which costs money. In unsupervised learning, there are none. Two more kinds sit in between: one uses only a few labels, and one makes its own.
Supervised learning is the easiest to check. You know the right answers for the test set, so you can count how often the model gets them. Its limit is the labels. Ten thousand labelled scans might cost a hospital weeks of doctors' time.
Unsupervised learning works on raw data, which is cheap and plentiful. It might split 10,000 shoppers into five groups. But it cannot tell you what each group means. A person has to look at the groups and decide, and some of them turn out to mean nothing.
Semi-supervised learning uses a few labelled rows and lots of unlabelled ones together. The labelled ones show what the answers look like. The rest show the shape of the data. With 200 scans marked by doctors and 20,000 unmarked, it often beats a model trained on the 200 alone.
Self-supervised learning makes its labels from the data itself. It hides part of an example and asks the model to guess it. The hidden part is the answer, so no person labels anything. BERT, from 2018, learned by guessing the 15% of word pieces hidden in its text. Chat models learn the same way, by guessing the next word.
People often think learning with no labels is better, since no one had to tag anything. It is not better, just different. If you know the answer you want and can get labels, use them. Learning from labels is nearly always the stronger choice.
The four kinds side by side
| Kind | Where the labels come from | Example |
|---|---|---|
| Supervised | People label every row | Spam or not spam |
| Unsupervised | No labels at all | Groups of shoppers |
| Semi-supervised | A few rows labelled, most not | 200 marked scans, 20,000 unmarked |
| Self-supervised | Hidden parts of the data itself | Guess the hidden word |
Which one fits
- 1
Supervised
When you know the answer you want and can pay for labels.
- 2
Unsupervised
When you have no labels and want to see what groups are there.
- 3
Semi-supervised
When labels cost a lot and raw data is cheap.
- 4
Self-supervised
When you have a vast pile of raw text or photos, and want a model to learn its shape.
Worked example
Emma learns from her orders
Emma has 50,000 past orders. Only 2,000 carry a tag saying whether the buyer came back.
She trains a supervised model on the 2,000 to guess who returns. She also runs an unsupervised one on all 50,000, and it finds a group of night-time buyers no one had noticed.
The words for this lesson4
Check yourself
How does self-supervised learning get its labels?
Finished this lesson?
Marking it complete also marks its 4 glossary entries learned, which updates your roadmap and your saved page. 0 are already marked.