Questions: 10 Pass mark: 8 out of 10 Retries: allowed after reading feedback
Instructions
Choose the best answer. The goal is careful judgement, not enthusiasm for larger models.
Questions
1. What is the main reason to use deep learning?
A. It always beats classical machine learning. B. It can learn useful representations when the task, data, and evidence process support it. C. It removes the need for validation. D. It makes small datasets safe.
2. Which task should usually start with a simpler method?
A. Image classification with enough rights-cleared labelled images. B. A tiny spreadsheet with 120 rows and unclear labels. C. A transfer-learning exercise using approved sample images. D. A tensor-shape practice notebook.
3. What is a parameter?
A. A trainable number inside a model. B. A private password. C. A notebook title. D. A row in every dataset.
4. What does a loss function do?
A. Measures how wrong the model output is for the training objective. B. Guarantees that the model is fair. C. Stores the dataset license. D. Chooses the GPU.
5. Why is CPU-first a good default for this course?
A. CPU models are always more accurate. B. It keeps core learning small, accessible, and reproducible without a cloud account. C. PyTorch cannot use GPUs. D. It prevents all runtime variation.
6. What should every substantial model notebook include?
A. Only the final score. B. A hidden helper that does everything. C. Evidence such as shapes, training result, validation result, errors, and limits. D. A claim that the model is ready for production.
7. What is the safest rule for private data in this course?
A. Paste it only into local notebooks. B. Use it if the model is small. C. Do not use private or sensitive data in course notebooks. D. Use it after deleting column names.
8. What is a state_dict useful for?
A. Saving model or optimizer state in an inspectable way. B. Proving that a model is correct. C. Replacing evaluation. D. Downloading a dataset.
9. What does a dataset card record?
A. Only the best metric. B. Source, permission, classification, contents, gaps, and suitable uses. C. The GPU driver version only. D. The model architecture only.
10. A model trains well but fails on validation. What is a reasonable first response?
A. Trust the training score. B. Make the model larger immediately. C. Check overfitting, split quality, label quality, learning rate, and data suitability. D. Delete the validation set.
Answers and Feedback
1. B
Deep learning is useful when representation learning may help and the evidence process is strong enough.
2. B
Small unclear tabular data usually needs data review, baselines, and simpler models first.
3. A
Parameters are trainable values the model changes during learning.
4. A
Loss measures model error for the objective. It does not prove real-world safety.
5. B
CPU-first keeps the required path accessible and easier to debug.
6. C
Deep learning needs evidence, not only a final score.
7. C
The course uses safe, public, generated, or synthetic material only.
8. A
Saving state_dict is the preferred course habit for small model artifacts.
9. B
A dataset card keeps source, permission, gaps, and use boundaries visible.
10. C
Validation failure is a signal to diagnose, not to hide.
What to Review
- Missed 1 or 2: review Unit 1.
- Missed 3, 4, or 8: review the setup map.
- Missed 5: review CPU, GPU, and reproducibility.
- Missed 6, 9, or 10: review evidence habits.
- Missed 7: review dataset safety.