Case context: A product team trains a mobile app classifier using customer photos taken on phones. They then add images downloaded from the company website. The website images are sharper, larger, and usually centered more neatly than the phone photos. The model must now learn those extra differences as well.

Question: Apply the same reasoning to explain the capacity risk and identify which data distribution should be protected.

Sample answer: The risk is that the model will use part of its representational capacity to learn the website photos’ sharper resolution, larger size, and cleaner centering. Because these properties are specific to the website images and differ from the phone-photo data, less capacity may be left for the phone-photo distribution. The team should reserve capacity for the customer phone photos, since that is the distribution the system needs to perform well on.

Key points:

• Sharper resolution, larger size, and cleaner centering are source-specific properties.
• Learning them consumes representational capacity.
• That consumption can reduce performance on the phone-photo distribution.
• The target distribution is the customer phone photos.

Rubric: The response should name the website-specific properties, explain that learning them uses representational capacity, connect that to possible harm on phone-photo recognition, and identify the customer phone photos as the distribution of interest.