S5 ICT: AI and Machine Learning, metric arithmetic and model size
- 2026-11-20
S5 ICT: AI and Machine Learning, metric arithmetic and model size
This paper is arithmetic heavy: F1 from precision 0.8 and recall 0.889, steps per epoch, image feature counts, dense layer parameters, the false comfort of accuracy on imbalanced data, convolution output size and a gradient step.
Challenge: metric arithmetic and model size This paper concentrates on computation: precision 0.8 with recall 0.889 gives F1 near 0.842, 12 000 samples in batches of 64 give 188 steps, a 224 by 224 colour image has 150 528 features, 784 into 128 gives 100 352 weights, ninety per cent of 3 000 is 300 wrong, 990 negatives against 10 positives gives accuracy 0.99 with recall 0, a 3 by 3 kernel on 28 by 28 gives 26 by 26, and a weight of 2.0 with gradient 3.0 at learning rate 0.1 becomes 1.7.
01 Paper details (check answers afterwards)
| Item | Detail |
|---|---|
| Questions | 10 |
| Suggested time | 25 |
| Difficulty mix | Foundation 3 · Intermediate 3 · Challenge 3 · Error redo 1 |
| Chapter | Chapter: S5 ICT: Artificial Intelligence and Machine Learning |
| Level | Challenge (10 items) |
02 Questions (10 items)
- FoundationPrecision is 0.80 and recall is 0.889. What is the F1 score?
- FoundationWith 12 000 samples per epoch and a batch size of 64, how many steps are there?
- FoundationHow many input features does a 224 by 224 colour image have?
- IntermediateA fully connected layer of 128 units follows an input of 784. How many weights and biases are needed?
- IntermediateA model scores ninety per cent on 3 000 training samples. How many are wrong?
- IntermediateWhat does lowering the decision threshold do to precision and recall?
- ChallengeWith 990 negatives and 10 positives, what are accuracy and recall if every case is predicted negative?
- ChallengeA 28 by 28 image passes a 3 by 3 kernel with no padding. What is the output size?
- ChallengeA weight of 2.0 has gradient 3.0 and the learning rate is 0.1. What is the weight after one update?
- Error redoEach of 1 000 images gets four augmentations. How many times larger is the dataset and how many images are there?
03 Answers and one-line notes
| # | Answer | One-line note |
|---|---|---|
| 1 | F1 is about 0.842. | F1 is the harmonic mean of the two, so the weaker one pulls the score down. |
| 2 | 188 steps. | 12 000 divided by 64 is 187.5, and the short final batch still counts as one step. |
| 3 | 150 528. | 224 by 224 pixels times three colour channels. |
| 4 | 100 352 weights and 128 biases. | 784 times 128 is 100 352, and each unit carries one extra bias term. |
| 5 | Three hundred. | 3 000 times 0.1 is 300, and those errors deserve individual inspection. |
| 6 | Recall rises and precision falls. | More samples are labelled positive, so fewer are missed and more are wrongly flagged. |
| 7 | Accuracy 0.99 and recall 0. | On imbalanced data accuracy hides the fact that the model never finds the positive class. |
| 8 | 26 by 26. | 28 minus 3 plus 1 is 26, since the kernel slides one pixel at a time. |
| 9 | 1.7. | 2.0 minus 0.1 times 3.0 is 1.7, and the gradient sets the direction of the step. |
| 10 | Five times larger, 5 000 images in total. | The original plus four augmented copies grows the set without new collection. |
04 What to do after marking
- After marking, redo only the wrong items and write down which step went wrong (formula, unit, or misread number).
- Repeat the same item two days later; move to the next paper only when both attempts are correct.
- Parents only need the answer and note columns; there is no need to explain every item.
Want to compute precision and recall by hand? Book a TutorZone tutor and work through all ten items in 25 minutes.Tutor Registration →
Questions and answers are written originally by the TutorZone team and are not taken from any school, publisher or examination authority paper. Every value is checked by computer.



