Supervised, Unsupervised, Reinforcement
Machine learning splits into three. Every question the board asks about them is a little story — and every one of them is answered the same way: look at what the machine was given.
1The one question that decides it
Unlabelled data → unsupervised.
No data, only rewards and penalties → reinforcement.
Tap each of the three to see where it sits and what hangs below it.
Full lineage: AI Models → Learning Based → Machine Learning → Supervised
2Supervised learning
The dataset fed to the machine is labelled. The person training the machine knows the data — that is the only reason they are able to label it in the first place.
Supervised means there is a supervisor
Feature: the weight. Label: the currency.
3Unsupervised learning
The dataset is unlabelled. The data is random, and the person training the model may have no information about it at all. The model’s job is to find the relationships, patterns and trends hiding inside — to tell you what the data is about.
Learning to swim with nobody watching
4Reinforcement learning
Here the computer makes a series of decisions that try to maximise a reward, with no human stepping in and without being explicitly programmed to do the task. It learns by repeated trial and error.
Typical examples: parking a car and a humanoid learning to walk.
5All three in one table
| Given | Used to | |
|---|---|---|
| Supervised | Labelled data | Determine relationships through training |
| Unsupervised | Unlabelled data | Discover new patterns |
| Reinforcement | No dataset — only rewards and penalties | Learn by rewarding actions |
6Naming the model from a story
| The scenario | The answer, and why |
|---|---|
| A social media site names your friend in a photo, from an album of tagged photographs | Supervised — the tags are labels |
| An OTT platform recommends films from your watch history | Unsupervised — it evaluates features of what you liked; nobody labelled anything |
| A bank flags suspicious transactions, with 'suspicious' nowhere defined | Unsupervised — with no 'fraud' label, the model looks for transactions unlike the rest |
| An autonomous vehicle is rewarded for arriving safely and penalised for traffic violations | Reinforcement — rewards and penalties, not a dataset |
| A company predicts customer churn from a dataset of customers labelled churned or not | Supervised — the labels are right there in the question |
7From the board papers
▶Show the answer
Reinforcement Learning. There is no dataset of correctly parked cars to learn from. The machine tries, is penalised when it clips a kerb and rewarded when it lands in the bay, and improves by trial and error.
▶Show the answer
Supervised and unsupervised learning both need a dataset to work from — labelled in one case, unlabelled in the other — and you need a fairly good idea of that data beforehand. Reinforcement learning uses no dataset at all. The machine learns by trial and error from rewards and penalties, aiming to maximise the reward, so it can be used where the environment changes or where there is not enough data about the situations the machine will meet.
Reason (R): Unsupervised learning models work on unlabelled datasets, where the data fed into the machine is random and the person training the model may not have any prior information about it.Class X handbook
▶Show the answer
Both (A) and (R) are true, and (R) is the correct explanation of (A). There being no labels is what “without any guidance” means here.
8Check yourself
Which learning approach uses labelled data for training?
A social media platform groups users by their interests, with no predefined categories. This is:
Which of the following is FALSE about reinforcement learning?
An AI agent playing a game and learning from its rewards and penalties is an example of: