Transparency, Interpretability and Explainability CT-AI Practice Questions (Page 4)

Questions 16–20

1. 16foundation · easy

   Why might an organization choose a less interpretable model despite the availability of more interpretable alternatives?

   • ABecause less interpretable models are always faster at inference time
   • BBecause less interpretable models are always easier to maintain
   • CBecause the less interpretable model offers better predictive performance on complex tasks
   • DBecause less interpretable models require less training data
   Reveal Solution Discussion Select an answer first
2. 17foundation · easy

   Which of the following is a common technique used to provide explanations for AI model decisions?

   • ALIME (Local Interpretable Model-agnostic Explanations)
   • BCross-validation
   • CGradient descent
   • DData augmentation
   Reveal Solution Discussion Select an answer first
4. 18application · medium

   A credit card company uses a complex ensemble model to detect fraudulent transactions. The fraud investigation team needs to understand why a specific transaction was flagged to decide if it should be reviewed. The model risk management team needs to verify that the model's overall behavior aligns with regulatory expectations. Which pair of interpretability techniques should the company use?

   • AUse LIME for the fraud team to explain the specific flagged transaction, and use SHAP to provide a global feature importance summary for the risk management team.
   • BUse a decision tree for the fraud team to explain the specific transaction, and use LIME for the risk management team to get a global view.
   • CUse a global surrogate model, like a decision tree, to explain the specific transaction to the fraud team, and use the same model for the risk management team.
   • DUse SHAP for the fraud team to explain the specific transaction, and use a partial dependence plot for the risk management team to get a global view.
   Reveal Solution Discussion Select an answer first
5. 19foundation · easy

   What is a common trade-off between interpretability and performance in AI models?

   • AMore interpretable models always have higher predictive performance
   • BInterpretability and performance are completely independent and do not affect each other
   • CMore interpretable models often have lower predictive performance than complex black-box models
   • DInterpretability only affects the training time, not the predictive performance
   Reveal Solution Discussion Select an answer first
6. 20foundation · easy

   Why is transparency considered important for stakeholders of an AI-based system?

   • AIt allows stakeholders to understand the system's purpose, data, and limitations, fostering trust and accountability
   • BIt makes the system's source code available to all users for modification
   • CIt guarantees the system will have the highest possible prediction accuracy
   • DIt ensures the system is fully automated and requires no human oversight
   Reveal Solution Discussion Select an answer first