S2 ICT AI practice 3: gradient descent, log loss and sample size, ten items
- 2026-11-20
S2 ICT AI practice 3: gradient descent, log loss and sample size, ten items
Ten items step a weight twice, then compute log loss, a survey sample size and the parameter count of a dense network before judging an AUC of 0.5.
S2 ICT AI practice 3: ten items S2 ICT AI practice 3: ten challenge items
01 Paper details (check answers afterwards)
| Item | Detail |
|---|---|
| Questions | 10 |
| Suggested time | 25 |
| Difficulty mix | Foundation 4 · Intermediate 2 · Challenge 2 · Error redo 2 |
| Chapter | Chapter: S2 ICT: Introduction to AI and machine learning |
| Level | The items combine gradient descent with model evaluation metrics. |
02 Questions (10 items)
- FoundationA weight of 2 takes one step with learning rate 0.1 and gradient 3. What is the new weight?
- FoundationFrom 1.7 the next gradient is 1.5 with the same learning rate. What is the weight now?
- FoundationThe model gives the correct class a probability of 0.9. What is the log loss?
- FoundationThe model gives the correct class only 0.1 probability. What is the log loss?
- IntermediateFor a 95 per cent confidence interval with plus or minus 3 per cent error and worst-case proportion 0.5, what minimum sample size is needed?
- IntermediateOne epoch takes 120 seconds. How long do 50 epochs take?
- ChallengeA dense network has 784 inputs, 128 hidden units and 10 outputs. How many trainable parameters?
- ChallengeA model reports an AUC of 0.5. What does that say about its discrimination?
- Error redoError: a larger learning rate always converges faster. What is correct?
- Error redoError: more training epochs always give a better model. What is correct?
03 Answers and one-line notes
| # | Answer | One-line note |
|---|---|---|
| 1 | 1.7. | The update is 0.1 times 3 which is 0.3, and 2 minus 0.3 leaves 1.7. |
| 2 | 1.55. | The second update is 0.1 times 1.5 which is 0.15, and 1.7 minus 0.15 leaves 1.55. |
| 3 | About 0.1054. | Log loss is the negative natural log of the probability, so minus ln 0.9 gives 0.1054. |
| 4 | About 2.3026. | Minus ln 0.1 is 2.3026, so low confidence is punished far more heavily. |
| 5 | About 1,068 cases. | 1.96 squared times 0.25 divided by 0.03 squared gives 1,067.11, rounded up to 1,068. |
| 6 | 6,000 seconds, which is 100 minutes. | 120 times 50 is 6,000 seconds, and dividing by 60 gives 100 minutes. |
| 7 | 101,770. | The first layer contributes 784 times 128 plus 128 biases and the second 128 times 10 plus 10 biases, giving 101,770. |
| 8 | It is no better than random guessing. | An AUC of 0.5 means positives and negatives are ranked correctly only half the time, while 0.9 would be good. |
| 9 | Too large a rate oscillates or diverges, so stability and speed must be balanced. | An oversized step can jump across the minimum so the loss rises instead of falling. |
| 10 | Too many epochs overfit, so stop when validation performance starts to fall. | Late in training the model starts memorising noise, so training should stop when validation error turns upward. |
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 more items on this chapter? The chapter page lists the other three papers in this set.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.



