Three kinds of machine learning
Machine learning comes in three main kinds: supervised learning (labelled examples), unsupervised learning (finding groups) and reinforcement learning (rewards).
📌 Key points
- Supervised learning learns from labelled examples: each input comes with the right answer, e.g. photos labelled "cat" or "dog".
- Two common supervised tasks: classification (choose a category, e.g. spam or not spam) and regression (predict a number, e.g. a house price).
- Unsupervised learning uses data with no labels. It finds structure by itself, such as groups of similar items. This is called clustering.
- Reinforcement learning uses trial and error. An agent takes actions and gets rewards for good results and penalties for bad ones, e.g. learning to play a game.
- Labels are often added by people, which can take a lot of time and money.
- To choose: labelled answers → supervised. No labels, want groups → unsupervised. Learning actions from rewards → reinforcement.
- Many real AI systems combine more than one kind.
📖 Key words
- Label
- The correct answer attached to a training example, e.g. "cat".
- Supervised learning
- Learning from examples that have labels (the right answers).
- Unsupervised learning
- Finding patterns or groups in data that has no labels.
- Clustering
- Putting similar items into groups.
- Reinforcement learning
- Learning by trial and error, using rewards and penalties.
- Agent
- In reinforcement learning, the program that takes actions and learns from rewards.
- Reward
- A score given for an action that helps reach the goal.
✏️ Worked example
Which kind of machine learning fits each task? (a) Spot broken bones using 5,000 X-rays marked by doctors. (b) Find groups of similar customers in shop data with no labels. (c) Learn to get through a maze game, scoring points for reaching the exit.
- (a) Each X-ray has the right answer (broken or not) from a doctor, so it is labelled data: supervised learning.
- (b) There are no labels. The computer must find groups by itself: unsupervised learning (clustering).
- (c) The program tries moves and gets points for good results: reinforcement learning.
- Answer: (a) supervised, (b) unsupervised, (c) reinforcement.
Ask yourself: are there labels? Labels → supervised. No labels → unsupervised. Rewards for actions → reinforcement.
Test yourself8 questions from a bank of 19. A new mix every time.
Start the quiz →