A clinic has only a few hundred labeled X-ray images for a model that flags a rare condition. Which technique creates more varied training examples without collecting new scans?
Show answer and explanation
Answer: B
Data augmentation builds new training examples by transforming existing ones. For images that means flips, rotations, crops and brightness shifts; for text it can mean synonym swaps or back-translation. The model sees more variation and generalizes better from a small labeled set. Adding layers increases capacity and tends to overfit when data is scarce. Larger instances speed up training but add no information, and lowering resolution discards detail without creating examples.
Why the other options are wrong
- A. More layers give the model more capacity, which worsens overfitting when labeled examples are scarce.
- C. Bigger instances shorten training time but do not add any new information to the dataset.
- D. Reducing resolution removes diagnostic detail and does not produce any additional training examples.
Exam tip. Augmentation stretches scarce labeled data; it does not replace collecting representative data.