AI Capabilities, Data Dependencies, and Failure Modes CAIPM Practice Questions (Page 7)

Questions 31–35

1. 31expert · hard

   An insurance company deploys a chatbot to handle customer claims. The chatbot was trained on historical chat logs. After deployment, customers complain that the chatbot gives incorrect answers for new policy types introduced after the training data was collected. The company has a limited budget and needs to fix the issue quickly. What is the best approach?

   • ACollect a large dataset of new policy interactions and retrain the model from scratch.
   • BReplace the chatbot with a rule-based system that has explicit rules for all policy types.
   • CImmediately retrain the chatbot on the original training data with more epochs to improve memorization.
   • DImplement a continuous monitoring system to detect when the chatbot's accuracy drops, and schedule periodic retraining with new chat logs.
   Reveal Solution Discussion Select an answer first
2. 32expert · hard

   A manufacturing company deploys a computer vision system to detect defects on an assembly line. The system was trained on images of products under consistent lighting. After a few months, the lighting conditions change due to new skylights, and the system starts missing defects. The company has limited labeled data from the new lighting conditions. Which combination of actions best addresses the failure mode while minimizing labeling effort?

   • AReplace the computer vision system with a rule-based system that uses fixed thresholds on pixel values.
   • BUse transfer learning to adapt the existing model with a small set of new labeled images, and monitor performance continuously.
   • CAdjust the model's decision threshold to be more sensitive to defects without retraining.
   • DCollect a large new dataset from the new lighting and retrain the model from scratch.
   Reveal Solution Discussion Select an answer first
4. 33expert · hard

   An e-commerce company wants to deploy a recommendation system for new users who have no purchase history. The company has a large dataset of user purchase histories, but new users have no data. Which approach is most likely to provide reasonable recommendations for new users?

   • AUse a hybrid recommender that combines content-based filtering with a popularity-based fallback for new users.
   • BUse collaborative filtering that relies on similar users' purchase histories.
   • CTrain a deep learning model on the purchase history and use it for new users.
   • DAsk new users to rate a set of items before providing recommendations.
   Reveal Solution Discussion Select an answer first
5. 34foundation · easy

   Which data characteristic is most critical for training a deep learning model that must recognize rare fraud patterns?

   • AQuantity of data
   • BData veracity
   • CData variety
   • DData velocity
   Reveal Solution Discussion Select an answer first
6. 35application · medium

   A financial services firm uses an AI model to approve small business loans. The model was trained on historical data that under-represented minority-owned businesses, and it now denies loans to those businesses at a higher rate. The firm wants to mitigate this bias while maintaining regulatory compliance. Which approach is most effective?

   • AIncrease the model's complexity to capture more subtle patterns in the data.
   • BRetrain the model on the same data but with a different random seed.
   • CRe-balance the training data to include more examples of minority-owned businesses and monitor the model's approval rates across groups.
   • DRemove all demographic features from the model to make it 'blind' to protected attributes.
   Reveal Solution Discussion Select an answer first