Pre-Trained Models and Transfer Learning CT-AI Practice Questions (Page 5)

Questions 21–25

1. 21application · medium

   A team has a pre-trained image classification model. They want to use it for a new task of classifying satellite images into land-use categories. They have a small labeled dataset. They decide to freeze the early layers of the pre-trained model and only train the later layers. What is the primary reason for freezing the early layers?

   • ATo speed up training by reducing the amount of data needed for the later layers.
   • BTo make the model more robust to adversarial attacks.
   • CTo reduce the risk of overfitting on the small dataset by keeping the general feature extractors unchanged.
   • DTo ensure the model does not forget the original task it was trained on.
   Reveal Solution Discussion Select an answer first
2. 22application · medium

   A company has a model trained to classify product reviews into positive/negative. They now want to classify reviews into a finer-grained set of categories (e.g., 'positive about price', 'negative about quality'). The source and target tasks are different but related, and both have labeled data. Which type of transfer learning is being applied?

   • AInductive transfer learning
   • BZero-shot transfer learning
   • CUnsupervised transfer learning
   • DTransductive transfer learning
   Reveal Solution Discussion Select an answer first
4. 23expert · hard

   A team has a model trained to classify images of animals. They want to use it to classify images of vehicles. They have labeled data for vehicles. Which type of transfer learning is this, and what is the primary consideration?

   • ATransductive transfer learning; the source and target domains are different, so the model must be retrained from scratch.
   • BInductive transfer learning; the source and target tasks are different but related, so the model can be fine-tuned on vehicle data.
   • CZero-shot transfer learning; the model can classify vehicles without any fine-tuning because it has general image knowledge.
   • DUnsupervised transfer learning; the model can be used without labeled data because it already knows how to classify images.
   Reveal Solution Discussion Select an answer first
5. 24foundation · easy

   What is the primary purpose of a pre-trained model in AI development?

   • ATo provide a rule-based system that can be easily interpreted by humans
   • BTo store data in a structured format for easy retrieval during inference
   • CTo serve as a starting point for new tasks, leveraging knowledge learned from a large dataset
   • DTo replace the need for any additional training or fine-tuning on new tasks
   Reveal Solution Discussion Select an answer first
6. 25foundation · easy

   What is 'catastrophic forgetting' in the context of transfer learning?

   • AWhen the model is unable to learn the target task at all
   • BWhen the model's performance degrades due to hardware failures
   • CWhen the model performs poorly on the target task due to overfitting
   • DWhen the model loses the knowledge it had from the source task after being fine-tuned on the target task
   Reveal Solution Discussion Select an answer first