Training and Testing a Model

1The six steps

1. Collect data. The class takes 30 photos of mangoes and 30 of avocados: ripe and unripe, big and small, on different backgrounds.
2. Label. They make two labels, mango and avocado, and put each photo under the right label.
3. Train. They click "Train". The computer studies the labelled photos and builds a model. This is called training.
4. Test. They show the model new photos it has never seen. For each one, the model gives its answer and how sure it is, for example "mango, 85% sure". This is called testing.
5. Improve. Where the model was wrong, they add more or better examples, fix any wrong labels, and train again.
6. Use. When it works well, they use the model, for example in a Scratch game that sorts fruit.

2Why test with NEW photos?

If you test a model with the same photos it trained on, it is like an exam that uses exactly the questions you memorised: you look clever, but it proves nothing. New photos show whether the model really learned what mangoes and avocados look like.

3How good is the model? Work out the accuracy

Accuracy tells you what part of the test the model got right, as a percentage:

accuracy = number correct ÷ number tested × 100

Example 1: The model is tested on 20 new photos and gets 17 right. 17 ÷ 20 = 0.85, and 0.85 × 100 = 85%.

Example 2: Another model gets 45 right out of 50. 45 ÷ 50 = 0.9, so the accuracy is 90%.

4When a model learns the wrong thing

One group photographed all their avocados on a green banana leaf and all their mangoes on a white plate. Their model scored well at first. Then they tested a mango lying on a banana leaf, and the model said "avocado"! The model had learned the background, not the fruit.

To fix it, they took new photos of both fruits on many different backgrounds and trained again. When a model gives unfair or one-sided answers because of the data it learned from, we say it has a bias. In real life, bias in AI can be serious: some models used to choose job applicants or to recognise faces have been criticised for treating some people unfairly.

Try it yourself

1. Put these steps in order: test, collect data, train, label.
2. A model gets 36 right out of 40 test photos. What is its accuracy?
3. Why must test photos be new photos?
4. A model calls every fruit on a banana leaf an "avocado". What went wrong, and how can you fix it?
5. Give one way to improve a model that makes many mistakes.

Answers

1. Collect data, label, train, test.
2. 36 ÷ 40 = 0.9, so 90%.
3. Because a model can do well on photos it has already seen without really learning; new photos show whether it learned the fruit.
4. It learned the background instead of the fruit. Take photos of both fruits on many different backgrounds and train again.
5. Any of: add more examples, add more varied examples, balance the numbers of each label, fix wrong labels, then train again.