中文（繁體）
中文（简体）
English
S2 ICT | Artificial Intelligence and Machine Learning lesson-prep guide (for tutors) | training and test sets, accuracy, bias
Every student has heard of artificial intelligence, but ask whether a model that is right nine times out of ten should be trusted, and most answer that nine out of ten is good enough. This lesson gives them three numbers to judge it: how much training data, how much test data, and which class it gets wrong.
Objectives — By the end students can: (1) explain why 200 images are split into 160 for training (eighty per cent) and 40 for testing (twenty per cent), and why the test set must never be used for tuning; (2) compute accuracy, so 36 correct out of 40 is 90%, and compute an all-cat baseline of only 60%, showing the ninety per cent is not luck; (3) use training at 95% (152 of 160) against testing at 90% (36 of 40), a five point gap, to explain overfitting, and name one concrete source of bias.
01 Prerequisites (3-minute check)
If any of these fails, fix it before the lesson starts.
| Skill | Quick check | Pass mark |
|---|---|---|
| Percentages | What is 36 out of 40 as a percentage? | 90%, because 36 divided by 40 is 0.9 |
| Ratio | Split 200 images eighty twenty. How many each side? | 160 for training, 40 for testing |
| Percentage points | How many percentage points apart are 95% and 90%? | Five points, not five per cent |
02 Lesson flow (45 minutes)
| Time | Stage | Tutor does | Student does |
|---|---|---|---|
| 0-5 min | Estimate first | Write “a model is right nine times out of ten, do you trust it”, then ask what happens if it answers one class every time | Hands go up, then doubt appears |
| 5-13 min | Split the data | Show 200 images becoming 160 and 40, and state that the test set stays locked | Can state the 80/20 split and the tuning rule |
| 13-22 min | Accuracy and baseline | Compute 36 over 40 as 90%, then the all-cat baseline of 24 over 40 as 60% | Both percentages right, gap of 30 points named |
| 22-32 min | Overfitting and bias | Compare 152 over 160 as 95% with the 90% test result and ask why | Can say “it memorised the training set” and “the data was not varied” |
| 32-40 min | Four question types | One question each on splitting, accuracy, baseline and bias | All four written with steps |
| 40-45 min | Exit check | Collect three answers and mark them, name the act to reteach | Hand in three answers |
03 Teaching script (read it out)
Scene 1: Act 1: training data is the textbook, test data is the exam
“We have 200 animal photos: 120 cats and 80 dogs. Take eighty per cent, 160, as the training set and the remaining twenty per cent, 40, as the test set. Why split at all? Because you need to know whether the model memorised the answers. Training data is the textbook, test data is the exam paper, and you may not look at the exam answers while tuning. The commonest mistake is tuning against the test score until it looks good and then reporting that score: that is reading the answers before the exam. Rule: train to learn, test to check, and use the test set once.”
Train to learn, test to check, and tuning on the test set is cheating.
Scene 2: Act 2: accuracy is correct over total, and it needs a baseline
“The test set of 40 holds 24 cats and 16 dogs. The model is right on 36, so accuracy is 36 divided by 40, which is 0.9, ninety per cent. The four errors are two cats called dogs and two dogs called cats. Is ninety per cent good? You need a baseline. A lazy model that always answers cat gets 24 right, which is 24 divided by 40, sixty per cent. So ninety beats sixty by thirty percentage points, which is real skill. The commonest misunderstanding is reporting accuracy without the class mix: if 39 of the 40 test images were cats, always answering cat would score 97.5%, better than our model. Report the class balance with the accuracy.”
Accuracy is correct over total, and it must be compared with the all-one-class baseline.
Scene 3: Act 3: training 95% against testing 90% is a gap about memorising
“On the 160 training images the model is right on 152, ninety-five per cent; on the test set, ninety per cent. The gap is five percentage points. If it had really learned pointed ears and round faces, both numbers would be close. The wider the gap, the more it looks like memorising, which is overfitting. Where does bias come from? Suppose all 120 training cats are white; the model may learn that white means cat and fail on the first black cat. So bias is not a broken model, it is data that is not varied enough. Three fixes: add photos in other colours, check the labels for mistakes, and shuffle before splitting. Adding layers does not fix biased data.”
A wide train-test gap means memorising, and bias comes from data that is not varied.
04 Question types and worked steps
| Type | Example | Steps | Answer |
|---|---|---|---|
| Splitting data | Split 300 images eighty twenty. How many each side? | (1) Multiply by 0.8 (2) multiply by 0.2 (3) add to check the total (4) give the unit | 240 for training and 60 for testing |
| Accuracy | A model is right on 43 of 50 test images. What is its accuracy? | (1) Correct over total (2) write the decimal (3) convert to a percentage (4) check it sits between 0 and 100% | 43 divided by 50 is 0.86, so 86% |
| Baseline | Of 40 test images, 24 are cats. What does always answering cat score? | (1) Find the largest class (2) divide it by the total (3) convert to a percentage (4) compare with the model | 24 over 40 is 60%, so the model at 90% is thirty points better |
| Bias analysis | All training cats are white and most black cats are misread. Why? | (1) Name the data bias (2) say the model learned colour (3) propose a fix (4) do not answer “the model is broken” | Data bias: colour was learned as the label; add varied colours and check the labels |
Sanity check (teach this): Sanity checks: training plus test must equal the total, 160 plus 40 is 200; accuracy must sit between 0 and 1, so writing 90 means the conversion was skipped; a baseline cannot exceed 100%, and a 97.5% baseline signals a badly unbalanced test set and must be pointed out; training accuracy is normally above test accuracy, so a much higher test score usually means the test set is tiny, where one error in five images moves the score by twenty points.
05 Independent practice (3 questions)
| Question | Answer | What to watch |
|---|---|---|
| Split 200 images eighty twenty. How many each side? | 160 for training and 40 for testing | Saying eighty per cent for the test set is the usual slip |
| A model is right on 36 of 40 test images. What is its accuracy? | 36 divided by 40 is 0.9, so 90% | Write the per cent sign, not just 0.9 |
| Training is 95% and testing is 90%. What does the five point gap suggest? | Likely overfitting, so more varied data or tighter validation is needed | Answering “the model is accurate” misses it |
06 Common mistakes and what to say
| Mistake | Student writes | Say this |
|---|---|---|
| Tuning on the test set | “I kept tuning until the test score hit 95%.” | “That test set is now spent. Carve a validation slice out of the training data to tune, and open the test set once at the end.” |
| Reporting accuracy without the class mix | “Ninety per cent, that is great.” | “Compare it with the baseline. If most test images are one class, always answering that class scores ninety-five, so your ninety is worse than guessing.” |
| Writing a percentage as a whole number | “The accuracy is 90.” | “90 is a count, not a percentage. 36 over 40 is 0.9, written as 90 per cent.” |
| Adding layers to fix bias | “The model is too shallow, add more layers.” | “Biased data stays wrong however deep you go. An all-white training set just produces confident mistakes, so add other colours first.” |
07 Homework (paper, section and question numbers)
- Chapter practice: split 200 images and write both numbers — AI chapter, practice question 1
- Chapter practice: compute two accuracies and compare them with the baseline — AI chapter, practice question 2
- Extension: find a classifier on your phone, such as face grouping or spam filtering, and describe one way it could learn the wrong thing — AI chapter, practice question 3
08 Exit check (last 5 minutes)
| Question | Answer | Decision |
|---|---|---|
| Why may the test set not be used for tuning? | It is reading the answers early, so the score is inflated and meaningless | Missing this means Act 1 again |
| 36 correct out of 40. What is the accuracy? | 90%, because 36 divided by 40 is 0.9 | Answering 36 means Act 2 again |
| What does a 95% training and 90% test gap suggest? | Possible overfitting, that is, memorising the training set | Answering “the model is accurate” means Act 3 again |
09 Tutor FAQ
Should network structure be taught?
Not in S2. Five words carry the lesson: data, training, testing, accuracy and bias. Whether a model can be trusted is a more useful skill than drawing neurons.
Does this need computers?
No. Two hundred images can be two hundred cards, or simply the class register: eighty per cent train, twenty per cent test. Hand-computed percentages meet the objective.
Should chatbots be discussed?
A sentence is enough. The point is that such systems also fail, and the failures follow the data; a product demo turns the lesson into a tour.
Find a tutor — freeTutor Registration →
A TutorZone lesson-prep guide. Written by the TutorZone team; questions and lesson flow are original and are not taken from any school or publisher material.