Case context: A team is building a model that decides whether rental bikes in customer-submitted photos need repair. Their labeled target set comes from a mobile repair app, and they also have many labeled bike photos from a public website. In both sources, the label depends only on the image itself, not on where it was captured or uploaded.

Question: Should the website photos be added to training under the idea of a consistent auxiliary data source? State the condition that makes the sources compatible and describe the expected downside and upside.

Sample answer: Yes. The website photos are compatible because the same mapping from image x to label y applies no matter whether the photo came from the app or the website; the source identity is not needed to predict the label. The main downside is extra compute or data-processing effort, while the upside can be a large gain from having more training examples.

Key points:

• The auxiliary photos are consistent if one predictor maps x to y without needing source identity.
• The origin of the photo does not have to be known to assign the label.
• The main cost is computational, and the likely benefit is more data.

Rubric: Grading Rubric:

• Explains that the website photos should be included because they are a consistent auxiliary source.
• Explains the mapping condition: a function from x to y works without source information.
• Notes the main downside is added compute/data handling, with a potential benefit from more examples.