Case context: A team is building a Siamese cat detector app. They train a single end-to-end classifier directly on raw images with 0/1 labels indicating whether the image contains a Siamese cat, but they only have a small labeled dataset and the model performs poorly.

Question: Based on the pipeline decomposition concept, what should the team do differently to reduce their data requirements and improve performance?

Sample answer: The team should decompose the end-to-end classifier into a pipeline: first, a cat detector that identifies whether a cat is present in the image, and second, a cat breed classifier that determines whether the detected cat is a Siamese cat. Each component solves a simpler sub-task than the combined end-to-end problem, so each requires significantly less data to train effectively than the single end-to-end classifier.

Key points:

• Single end-to-end classifier requires large amounts of data for the combined task
• Decomposing into a cat detector and cat breed classifier creates simpler sub-tasks
• Each pipeline component requires significantly less data than the end-to-end model
• This decomposition directly addresses the team's limited data problem

Rubric: Full credit for recommending decomposition into a cat detector and cat breed classifier and explaining that each simpler component needs less data. Partial credit for recommending decomposition without explaining the data benefit.