Knowledge check

Quiz

In Progress

Knowledge check

Evaluate your understanding of this unit by completing the knowledge check below.

Action item: Knowledge check

Work through each question. Correct answers and feedback mirror the Multiverse assessment.

Question 1 of 10

1. You are fine-tuning a neural network for image classification. You increase the learning rate slightly, and the loss oscillates instead of decreasing. What is the most likely cause?

A. The model has too few layers.
B. The learning rate is too low.
C. The learning rate is too high.
D. The batch size is too small.

Correct Answer: C

The best choice is (C): The learning rate is too high.

2. Which hyperparameter tuning method is most computationally efficient when evaluating expensive models?

A. Grid search
B. Random search
C. Bayesian optimization.
D. Manual tuning

Correct Answer: C

The best choice is (C): Bayesian optimization.

3. You are training a random forest classifier and notice diminishing returns after increasing the number of estimators from 100 to 300. What’s the best next step?

A. Reduce the number of estimators to balance performance and cost.
B. Continue increasing the number of estimators to 1,000.
C. Decrease the learning rate.
D. Replace the model with a single decision tree.

Correct Answer: A

The best choice is (A): Reduce the number of estimators to balance performance and cost.

4. In boosting algorithms like XGBoost, what is the main purpose of sequentially training new learners?

A. To randomly diversify the training subsets.
B. To correct errors made by previous learners.
C. To reduce training time by parallelising learning.
D. To increase interpretability.

Correct Answer: B

The best choice is (B): To correct errors made by previous learners.

5. A single decision tree model on a credit scoring dataset performs well on training but poorly on unseen data. Which ensemble technique would most likely help?

A. Boosting using XGBoost.
B. Stacking with logistic regression.
C. Dropout regularisation.
D. Bagging using Random Forest.

Correct Answer: D

The best choice is (D): Bagging using Random Forest.

6. You deploy a model that predicts a 70% chance of default, but historically only 50% of such cases default. What does this suggest?

A. The model is underconfident.
B. The model is overconfident.
C. The model is well-calibrated.
D. The model has a high bias.

Correct Answer: B

The best choice is (B): The model is overconfident.

7. Which technique would best correct an SVM model’s poorly calibrated probability outputs?

A. Platt scaling
B. Early stopping
C. Random search
D. Bagging

Correct Answer: A

The best choice is (A): Platt scaling

8. You’re building a loan default model where accurate probability estimates are critical for setting risk thresholds. What should you prioritise after tuning and ensembling?

A. Increasing the model’s accuracy on training data.
B. Ensuring model calibration aligns predicted probabilities with actual outcomes.
C. Using a more complex ensemble.
D. Adding regularisation to reduce overfitting.

Correct Answer: B

The best choice is (B): Ensuring model calibration aligns predicted probabilities with actual outcomes.

9. You’re short on GPU resources but need to optimise a neural network for AUC. Which tuning approach is most practical?

A. Large-scale grid search over multiple parameters.
B. Bayesian optimisation with 200 iterations.
C. Random search with early stopping.
D. Automated neural architecture search (NAS).

Correct Answer: C

The best choice is (C): Random search with early stopping.

10. After tuning and calibration, your model performs well but takes too long to train. What’s the most effective next step?

A. Reduce model complexity or ensemble size while monitoring performance.
B. Remove calibration to speed up training.
C. Disable regularisation entirely.
D. Double the learning rate.

Correct Answer: A

The best choice is (A): Reduce model complexity or ensemble size while monitoring performance.