Demonstrate the Ability to Train and Inference CERTIFIED-AI-EXPERT Practice Questions (Page 3)

Questions 11–15

1. 11foundation · easy

   When using a trained image classification model for inference, what is the purpose of resizing and normalizing the input image before feeding it to the model?

   • ATo match the input format expected by the model
   • BTo convert the image to a different color space for storage
   • CTo improve the model's training accuracy
   • DTo reduce the number of model parameters
   Reveal Solution Discussion Select an answer first
2. 12application · medium

   A team is preparing a dataset for training a language model. The dataset contains text with many duplicate entries and inconsistent formatting. Which of the following steps is most important in the data preparation phase?

   • AImmediately start training to save time, as the model can handle noisy data.
   • BRandomly sample a subset of the data to reduce training time.
   • CConvert all text to a single character encoding, such as ASCII.
   • DDeduplicate the data and standardize the text formatting (e.g., lowercasing, removing extra spaces).
   Reveal Solution Discussion Select an answer first
4. 13application · medium

   A model that detects spam emails is deployed. For a new email, the model outputs a probability of 0.9. The company has set a threshold of 0.5 for classifying an email as spam. What is the classification of this email?

   • ASpam, because the probability exceeds the threshold.
   • BThe email is spam with 90% confidence, but the threshold should be adjusted.
   • CThe model is uncertain, so the email should be manually reviewed.
   • DNot spam, because the probability is not high enough.
   Reveal Solution Discussion Select an answer first
5. 14expert · hard

   A company is deciding whether to train a model in-house or use a pre-trained model and fine-tune it. The training data is sensitive and cannot leave the company's premises. Which of the following is the most important consideration?

   • AUsing a pre-trained model is always faster and cheaper, so it should be chosen regardless of data sensitivity.
   • BTraining in-house ensures data stays within the company's controlled environment, but may require significant compute resources.
   • CTraining in-house is always more accurate than using a pre-trained model.
   • DFine-tuning a pre-trained model on-premises is impossible because pre-trained models are only available in the cloud.
   Reveal Solution Discussion Select an answer first
6. 15expert · hard

   A model serving team must deploy a model that has variable inference traffic, with occasional spikes. They need to minimize cost while ensuring availability during spikes. Which of the following deployment strategies is most appropriate?

   • AProvision a fixed number of always-on instances to handle the peak load.
   • BUse a batch processing system that queues requests and processes them periodically.
   • CDeploy on a single powerful GPU instance and rely on it to handle spikes.
   • DUse a serverless inference platform that automatically scales with demand and charges per invocation.
   Reveal Solution Discussion Select an answer first