EXP.AI.01 Explore · AI Lab
How machines learn from examples
You should be able to: explain machine learning as “learn from many examples,” and run a paper activity that behaves like a tiny classifier.
After GES: B7.4.4.1.1 names ML, ANN, VR, and AR. This lesson shows how learning from data differs from a fixed calculator program.
1. Learn
A normal program is a recipe you write in full: if mark ≥ 50 then pass. The computer does not invent the 50.
Machine learning (ML) is different. You give many examples (photos of maize leaves: healthy vs diseased). The system finds patterns. Later it guesses for a new leaf it has not seen. That guess can be wrong — especially if your photos were only taken in one region or one season.
- Training data — the examples you showed it.
- Model — the pattern it stored (not a person in the box).
- Inference — using the model on a new case.
This UltraHost plan cannot run ChatGPT or a local AI model. You will learn the idea with paper, then you may try a teacher-approved public tool later (Teachable Machine in a lab, if the school allows it).
Rule: ML is not magic and not always fair. If the data is mostly one kind of example, the model will be weak on everything else. That is bias.
2. See
Paper activity (5 minutes): cut or draw 20 cards. Mark 12 as “tro-tro” (many windows, long) and 8 as “private car”. A partner, who did not see your marks, only sees crude features you announce: “long / short”, “high / low”. They write a rule. Then show 4 new drawings. That rule is a tiny model. Count how many it gets wrong.
3. Practice
-
In one sentence, how is ML different from a calculator program?
Show answer
A calculator follows fixed rules you typed. ML changes its behaviour after seeing many examples.
-
A clinic trains a model on X-rays from adults in Accra only, then uses it on children in a village. What is the risk?
Show answer
The training data does not match the new patients (age, equipment, disease mix). The model may give dangerous wrong answers. That is a data / bias problem.
-
Match: training data, model, inference — (a) a new maize photo today; (b) last season’s 2,000 labelled photos; (c) the stored pattern.
Show answer
(a) inference input (b) training data (c) model
-
Why does SURF-CLOUD not run a big AI model on this website?
Show answer
Shared hosting has about 1GB RAM and is built for PHP pages, not GPU models. We teach the ideas and use external tools when needed.
4. Project
Write a half-page: one Ghana use of ML (farm, health, MoMo fraud alerts, translation) and one way it could fail if the data is poor. Bring your paper “tro-tro vs car” score (right / 4).
5. Check
- Emerging — “AI is a smart computer” only
- Developing — names examples but not training vs inference
- Approaching proficient — explains ML vs rules with a Ghana use
- Proficient — finishes the paper activity and states a bias risk
- Highly proficient — leads the card game for a small group
SURF-CLOUD Explore track · beyond GES Computing. Next: EXP.AI.02.