Compare the Differences Between Model Building and Fine-Tuning Models CERTIFIED-AI-EXPERT Practice Questions (Page 2)

Questions 6–10

1. 6application · medium

   A team is developing a model to detect fraudulent transactions in a highly specialized financial system. They have access to a massive dataset of 100 million labeled transactions. The patterns of fraud are very unique to their system and not similar to any public dataset. What is the most appropriate approach?

   • ATrain a model from scratch on the 100 million transactions.
   • BFine-tune a pre-trained anomaly detection model on the 100 million transactions.
   • CFine-tune a pre-trained model on a small subset of 10,000 transactions to save time.
   • DUse a pre-trained model as a fixed feature extractor and train a shallow classifier on the extracted features.
   Reveal Solution Discussion Select an answer first
2. 7foundation · easy

   In terms of data labeling, what is a typical advantage of fine-tuning over building from scratch?

   • AFine-tuning can often achieve good performance with fewer labeled examples.
   • BFine-tuning requires more labeled data because it must learn from scratch.
   • CFine-tuning requires labeled data for both the pre-training and fine-tuning phases.
   • DFine-tuning requires no labeled data at all.
   Reveal Solution Discussion Select an answer first
4. 8foundation · easy

   In which scenario is building a model from scratch most appropriate?

   • AWhen the target task is highly specific and no pre-trained model exists for a similar domain.
   • BWhen you have a very small dataset and limited computational resources.
   • CWhen you need to deploy the model quickly and cheaply.
   • DWhen the task is a common image classification problem with many available pre-trained models.
   Reveal Solution Discussion Select an answer first
5. 9application · medium

   A startup is building a recommendation system for a niche e-commerce platform. They have 1 million user-item interactions. They have a limited compute budget and need to deploy the model quickly. The interactions are sparse and the item catalog is unique. What is the most appropriate approach?

   • AUse a pre-trained model as a fixed feature extractor and train a linear regression on the extracted features.
   • BTrain a collaborative filtering model from scratch on the 1 million interactions.
   • CFine-tune a large deep learning model on the 1 million interactions.
   • DFine-tune a pre-trained recommendation model on the 1 million interactions.
   Reveal Solution Discussion Select an answer first
6. 10application · medium

   A company wants to build a model to predict customer churn for their subscription service. They have 2 million labeled customer records. The patterns of churn are fairly standard and similar to other industries. They have a limited budget and need the model in two weeks. What is the most appropriate approach?

   • AFine-tune a pre-trained tabular model on the 2 million records.
   • BUse a pre-trained model as a fixed feature extractor and train a linear classifier.
   • CTrain a gradient boosting model from scratch on the 2 million records.
   • DFine-tune a large deep learning model on the 2 million records.
   Reveal Solution Discussion Select an answer first