Question: Using the example of a cat detector and cat breed classifier, explain why decomposing an end-to-end classification task into pipeline components can reduce the amount of training data needed compared to training a single end-to-end classifier on 0/1 labels.

Sample answer: An end-to-end classifier must learn a single complex mapping from raw pixels directly to a 0/1 label, which implicitly requires learning to both detect a cat and distinguish breeds within one model. By decomposing this into a cat detector and a cat breed classifier, each component only needs to learn a simpler, more focused sub-task. Simpler tasks require fewer examples to learn well, so each pipeline component needs significantly less data than the harder combined end-to-end task.

Key points:

• End-to-end classifier learns one complex mapping directly from raw input to 0/1 label
• Pipeline decomposition splits this into simpler focused sub-tasks
• Cat detector and cat breed classifier are each easier to learn individually
• Simpler sub-tasks require significantly less data than the combined end-to-end task

Rubric: Full credit for explaining that decomposition creates simpler sub-tasks per component, that simpler tasks need less data, and that this results in less data required overall compared to the end-to-end classifier. Partial credit for a vague or incomplete comparison.