中文（繁體）
中文（简体）
English
S6 ICT: Artificial Intelligence and Machine Learning – Practice Paper 2 (Intermediate, 10 Questions)
S6 ICT: Artificial Intelligence and Machine Learning – Practice Paper 2 (Intermediate, 10 Questions)
Paper two moves up a level: 80/20 splitting, five-fold cross validation, a 33.5 percentage point train-test gap, the 90.0% trap of class imbalance, precision 80.0% with recall 60.0% into F1, batch counts and threshold tuning.
Quick answer At this level watch the gap and the denominator: a wide train-test gap means overfitting, and precision is measured over all positive predictions.
01 Paper details (check answers afterwards)
| Item | Detail |
|---|---|
| Questions | 10 questions |
| Suggested time | 25 minutes |
| Difficulty mix | Foundation 2 · Intermediate 3 · Challenge 3 · Error redo 2 |
| Chapter | Chapter: S6 ICT | Artificial Intelligence and Machine Learning |
| Level | Intermediate |
02 Questions (10 items)
- Foundation600 images are split 80/20 into a training set and a test set. How many images in each?
- FoundationThe same 600 images are used for five-fold cross validation. How large is each test fold, and how many training runs are there?
- IntermediateModel A reaches 98.5% on the training set and 65.0% on the test set. By how many percentage points do they differ?
- IntermediateA data set holds 900 non-target images and 100 target images. What is the accuracy of a classifier that labels everything as non-target?
- IntermediateA model produces 48 true positives and 12 false positives on the test set. What is its precision?
- ChallengeThe same model also produces 32 false negatives. What is its recall?
- ChallengeWith a precision of 80.0% and a recall of 60.0%, compute the F1 score to three decimal places.
- ChallengeThe training set holds 480 images and the batch size is 32. How many batches are there in one epoch?
- Error redoLowering the decision threshold from 0.5 to 0.3 raises the number of positive predictions from 100 to 130. How does this affect precision and recall?
- Error redoA student sees 98.5% on the training set and says the model is ready to use. What is being ignored?
03 Answers and one-line notes
| # | Answer | One-line note |
|---|---|---|
| 1 | 480 for training and 120 for testing. | 600 times 0.8 is 480, and 600 times 0.2 is 120. |
| 2 | 120 images per fold, five training runs. | 600 divided by 5 is 120, and each fold takes its turn as the test set, so five runs are needed. |
| 3 | 33.5 percentage points. | 98.5 minus 65.0 is 33.5, and a gap this wide is a classic sign of overfitting. |
| 4 | 90.0%. | 900 divided by 1000 is 0.90, so the accuracy is high even though the model is useless. |
| 5 | 80.0%. | 48 divided by (48+12) is 48 over 60, which is 0.80. |
| 6 | 60.0%. | 48 divided by (48+32) is 48 over 80, which is 0.60. |
| 7 | 0.686. | Two times 0.80 times 0.60 is 0.96, and dividing by 1.40 gives about 0.686. |
| 8 | 15 batches. | 480 divided by 32 is 15, and a final short batch still counts as one batch. |
| 9 | Recall usually rises while precision usually falls. | A lower threshold pushes more samples to positive, which cuts misses but adds false alarms. |
| 10 | Checking the training set alone hides overfitting. | The test set is only 65.0%, so the model has not learned patterns that generalise. |
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.
Find a tutor freeTutor 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.