Question: In a brief analytical response, explain why combining different training data sources, such as clinic photos and mobile-phone photos, can be risky for some learning systems but not for others. Compare older, less flexible methods with modern, highly flexible methods.

Sample answer: Older approaches, such as manually crafted visual features followed by a simple classifier, had limited ability to adapt to differences between sources, so mixing clinic photos and mobile-phone photos could reduce accuracy. Modern methods, such as large neural networks, are much more flexible and can absorb a wider range of variation, so combining those sources is usually far less risky.

Key points:

• Older methods with fixed feature design plus a simple classifier were more likely to suffer from mixed-source data
• Modern large neural networks are much less sensitive to this issue
• The main reason is the model's flexibility or capacity
• The example involves combining two different image sources

Rubric: Full credit requires naming both categories of algorithms (older hand-crafted-feature plus simple-classifier methods and modern large neural networks), describing the different risk levels for each, and explaining that the difference comes from model flexibility.