Question: A quality-control team wants software that labels a conveyor-belt photo as "accept" or "reject." They can either train one end-to-end model that maps the raw image directly to the binary label, or split the system into two parts: one module locates the circuit board in the photo, and a second module checks whether the solder joints are cracked. Explain why the two-part design can require less labeled data than the single end-to-end model.

Sample answer: A single end-to-end model has to learn several things at once. It must figure out where the relevant object is in the image and also learn which visual patterns indicate a defect, all while producing the final accept/reject decision. When the job is divided into two stages, each module handles a narrower problem. Finding the board is easier than learning the whole final decision, and checking the solder joints is easier once the relevant region has been isolated. Because each subtask is simpler, each one usually needs fewer labeled examples than the full combined task.

Key points:

• The end-to-end model must learn the full mapping from raw image to binary label
• The staged design breaks the problem into narrower subtasks
• Locating the board and checking solder joints are simpler than learning the entire accept/reject decision at once
• Simpler subtasks generally need fewer labeled examples

Rubric: Full credit for explaining that decomposition creates simpler subtasks, that simpler subtasks require less data, and that this can reduce the labeling burden compared with one end-to-end model. Partial credit for a generally correct but incomplete explanation.