AI Model Training, Testing, and Validation AAIR Practice Questions (Page 8)

Questions 36–40

1. 36foundation · easy

   What is an example of an adversarial input?

   • AAn image with subtle pixel perturbations that cause a classifier to misidentify the object
   • BA text input with correct grammar and spelling
   • CA data point that is exactly the same as a training example
   • DA completely new image that is similar to training examples
   Reveal Solution Discussion Select an answer first
2. 37expert · hard

   A university is deploying an AI system to predict student dropout risk. The model shows different false-positive rates for different ethnic groups. The administration wants to reduce this disparity without significantly reducing overall model performance. Which approach is most appropriate?

   • AApply a fairness constraint during training, such as equalized odds, and evaluate the trade-off with overall performance
   • BIncrease the model's complexity to capture more nuanced patterns
   • CRemove ethnicity as a feature and assume the model is now fair
   • DReport only the overall accuracy to avoid drawing attention to the disparity
   Reveal Solution Discussion Select an answer first
4. 38application · medium

   A bank is developing a credit-default model. The data science team reports that the model achieves 99.2% accuracy on the training set but only 87.4% accuracy on the held-out test set. The dataset is heavily imbalanced, with only 3% of applicants defaulting. Which action should the risk team recommend FIRST to properly evaluate this model?

   • ASplit the data again using a different random seed to get a more representative test set
   • BIncrease the size of the test set to 40% of the data to improve statistical confidence
   • CUse precision, recall, and F1-score on the test set instead of relying on accuracy
   • DReduce the model's complexity by removing features until training accuracy drops below 95%
   Reveal Solution Discussion Select an answer first
5. 39application · medium

   A medical diagnostics company is developing a model to screen for a rare disease that affects 0.1% of the population. The cost of a false negative (missing a case) is extremely high, while false positives are less harmful but still require follow-up testing. Which metric should the team prioritize when evaluating the model?

   • APrecision, because it measures the proportion of positive predictions that are correct
   • BRecall (sensitivity), because missing a case is the most costly outcome
   • CAccuracy, because it provides an overall measure of correctness
   • DF1-score, because it balances precision and recall
   Reveal Solution Discussion Select an answer first
6. 40foundation · easy

   Which situation is most likely to cause underfitting in a machine learning model?

   • ARemoving outliers from the training data
   • BUsing a model that is too simple to capture the underlying patterns in the data
   • CTraining the model for too many epochs on a large dataset
   • DIncluding too many irrelevant features in the model
   Reveal Solution Discussion Select an answer first