Case context: A team works with a game studio to generate 60,000 labeled images of bicycles. The rendering system can change lighting, viewpoint, and background, but it only uses 15 bicycle frame templates. A classifier trained on these synthetic images reaches very high training accuracy. However, when the model is tested on real photographs of bicycles taken on city streets, its accuracy drops sharply.

Question: What is the most likely reason for the poor test performance, and what does this show about the difference between having many training examples and having diverse training examples?

Sample answer: The most likely reason is overfitting to the small set of 15 bicycle templates. Although the training set is large, most images are variations of the same underlying shapes, so the model learns patterns that work well for those specific templates instead of features that generalize to bicycles in the real world. The street photos contain many more kinds of frames, accessories, camera angles, and backgrounds, so the learned features do not transfer well. This shows that a large number of examples is not enough by itself; the training data also needs enough diversity to match the kinds of inputs the model will face later.

Key points:

• Very high training accuracy with poor test accuracy suggests overfitting
• The main limitation is the small number of underlying templates, not the raw image count
• Real street photos have much broader variation than the synthetic renders
• Diversity matters more than quantity when the goal is generalization

Rubric: Full credit: identifies overfitting to the 15 templates, explains why many near-duplicate images are not enough, and connects the failure to the broader real-photo distribution. Partial credit: identifies overfitting but gives only a partial explanation of the diversity issue.