Introduction to AI

Lessons

Senior 1 · ICT

Introduction to AI

5 lessons

Lesson 5 of 5

Introduction to Machine Learning: Training and Testing Data

25 min

How do machines learn?

Most of the AI you met in this unit, from face unlock to spam filters, works because of machine learning (ML). The computer scientist Arthur Samuel introduced the term in 1959, while working on a program that learned to play draughts (checkers) better than he could. In this lesson you will see how machine learning differs from ordinary programming, why its data is split into training and testing sets, and how it affects daily life.

Traditional programming versus machine learning

Machine learning is a branch of AI in which a computer learns rules from examples, instead of a programmer writing every rule.

plaintext

Traditional programming:   data + rules    --> [ computer ] --> answers
Machine learning:          data + answers  --> [ computer ] --> rules (a model)

Code Explanation

In traditional programming, a person writes the rules (for example, "if the temperature is above 30 °C, show Hot") and the computer applies them to data.
In machine learning, the computer is given many examples, each with its correct answer (a label), and works out the rules itself. What it learns is called a model. The model is then used on new data.

Worked example: ripe or unripe coffee cherries?

A coffee cooperative in Huye wants a phone app that tells pickers whether a coffee cherry is ripe from its redness score (0 to 100), measured by the phone's camera.

Step 1: collect and label training data. Expert pickers score and label eight cherries:

Redness	20	35	45	50	70	75	85	90
Label	unripe	unripe	unripe	unripe	ripe	ripe	ripe	ripe

Step 2: train. The model looks for the boundary that separates the two labels. The highest unripe score is 50 and the lowest ripe score is 70, so it places the boundary halfway: (50 + 70) ÷ 2 = 60. The learned rule is: ripe if redness is 60 or more. Nobody wrote "60"; it came from the data.

Step 3: test. The team checks the model on five new cherries that it never saw in training, also labelled by experts:

Redness	Expert's label	Model says	Correct?
30	unripe	unripe	Yes
65	ripe	ripe	Yes
58	ripe	unripe	No
80	ripe	ripe	Yes
40	unripe	unripe	Yes

Accuracy = correct answers ÷ total tested × 100% = 4 ÷ 5 × 100% = 80%. The mistake at 58 shows that the model needs more training examples near the boundary.

The same steps written in Python (optional; you will learn to program later):

python

# Training data: (redness score, label given by expert pickers)
training = [(20, "unripe"), (35, "unripe"), (45, "unripe"), (50, "unripe"),
            (70, "ripe"), (75, "ripe"), (85, "ripe"), (90, "ripe")]

# Learning: find the edges of the two groups
highest_unripe = 0
lowest_ripe = 100
for redness, label in training:
    if label == "unripe" and redness > highest_unripe:
        highest_unripe = redness
    if label == "ripe" and redness < lowest_ripe:
        lowest_ripe = redness
boundary = (highest_unripe + lowest_ripe) / 2
print("Learned rule: ripe if redness >=", boundary)

# Testing: new cherries the model has never seen
testing = [(30, "unripe"), (65, "ripe"), (58, "ripe"), (80, "ripe"), (40, "unripe")]
correct = 0
for redness, label in testing:
    guess = "ripe" if redness >= boundary else "unripe"
    if guess == label:
        correct = correct + 1
print("Correct:", correct, "out of", len(testing))
print("Accuracy:", correct * 100 / len(testing), "%")

Code Explanation

Output: Learned rule: ripe if redness >= 60.0, then Correct: 4 out of 5 and Accuracy: 80.0 %. The loop over training is the learning; the loop over testing is the check.

Training data and testing data

Training data is the set of labelled examples the model learns from.
Testing data is a separate set of labelled examples, kept hidden during training, used to measure how well the model works on data it has never seen.
A common split is about 80% for training and 20% for testing. From 1,000 labelled photos, 800 would train the model and 200 would test it.

Why keep them separate? Testing a model on its own training examples is like giving students an exam made of the exact questions they practised: a high score would not prove they can answer new questions. Only new data shows whether the model has learned the pattern or just memorised the examples.

Good data is plentiful, correctly labelled, and varied: it includes every kind of case the model will meet (different lighting, different coffee varieties, different people). Poor or unbalanced data gives a poor or biased model.

Impact of machine learning in real life

Recommendations: video and music apps learn what you like. The risk: you may only see more of the same.
Spam filters learn from emails that users mark as spam.
Speech and translation: voice typing and translation improve as they learn from more recordings and texts, including Kinyarwanda.
Fraud detection: banks and mobile-money services learn each customer's normal pattern and flag unusual payments.
Health and farming: models learn to spot disease in scans and in leaf photos.
Weather forecasting uses models trained on years of past weather.

Tip

Remember the pipeline: collect → label → train → test → use. Train on one set, test on a different set, and measure accuracy = correct ÷ total × 100%. A model is only as good as its data.

✦

Practice

State one difference between traditional programming and machine learning.

Show solutionHide solution

Solution: In traditional programming, a person writes the rules and the computer applies them. In machine learning, the computer is given examples with answers and works out the rules (the model) itself.

A model is tested on 250 new photos and labels 225 correctly. What is its accuracy?

Show solutionHide solution

Solution: 225 ÷ 250 × 100% = 90%.

A team has 2,000 labelled images and uses an 80% and 20% split. How many images are used for training and how many for testing?

Show solutionHide solution

Solution: Training: 0.8 × 2,000 = 1,600. Testing: 0.2 × 2,000 = 400.

Gatete trains a model and tests it on the same images. It scores 100%. Why should the cooperative not trust this score?

Show solutionHide solution

Solution: The model has already seen those images, so it may have memorised them. Only testing on new, unseen data shows how it will perform in real use.

In the coffee example, a new training cherry with redness 55 is labelled ripe. Where would the new boundary be?

Show solutionHide solution

Solution: The lowest ripe score is now 55 and the highest unripe score is still 50, so the boundary is (50 + 55) ÷ 2 = 52.5. The cherry at 58 would now be classed correctly as ripe.

Summary

Machine learning: the computer learns rules (a model) from labelled examples instead of being given the rules.
Training data teaches the model; separate testing data checks it on unseen cases. A common split is 80% and 20%.
Accuracy = correct ÷ total × 100%. Plentiful, correct, varied data gives better, fairer models.
ML powers recommendations, spam filters, translation, fraud detection, medical and crop diagnosis and weather forecasts.

Sources and credits

Machine learning, Wikipedia contributors, Wikipedia, licensed under CC BY-SA 4.0. Adapted by Ganzaa.
Training, validation, and test data sets, Wikipedia contributors, Wikipedia, licensed under CC BY-SA 4.0. Adapted by Ganzaa.
Arthur Samuel (computer scientist), Wikipedia contributors, Wikipedia, licensed under CC BY-SA 4.0. Adapted by Ganzaa.

The coffee-cherry example, the Python program, the diagram and the practice items were written by Ganzaa.

This lesson adapts CC BY-SA material and is shared under CC BY-SA 4.0.

1How do machines learn?

2Traditional programming versus machine learning

Worked example: ripe or unripe coffee cherries?

3Training data and testing data

4Impact of machine learning in real life

Practice

How do machines learn?

Traditional programming versus machine learning

Training data and testing data

Impact of machine learning in real life