Questions: Data Augmentation Techniques

5 questions to test your understanding

Score: 0 / 5
Question 1 Multiple Choice

A researcher is training a model to classify handwritten digits (0–9). Which of the following augmentations would be INAPPROPRIATE because it could introduce incorrect training signal?

ARandomly adding Gaussian noise to pixel values
BRandomly scaling the image by 5–10%
CHorizontally flipping the image (left-right mirror)
DSlightly adjusting the brightness of the image
Question 2 Multiple Choice

A team applies aggressive data augmentation to a small medical imaging dataset — including random rotations up to 180°, horizontal/vertical flips, and random color channel inversions — and finds that model accuracy on the validation set *decreases* compared to training without augmentation. What is the most likely explanation?

AData augmentation always reduces accuracy in small datasets; it only helps with large datasets
BSome augmentations destroyed label semantics (e.g., flipping a chest X-ray changes its clinical interpretation), introducing incorrect training signal
CThe model was too small to learn from augmented data and needed more parameters
DAugmentation increases training time, causing the model to underfit due to insufficient epochs
Question 3 True / False

Data augmentation reduces overfitting by increasing the effective variety of the training set, making it harder for the model to memorize specific examples.

TTrue
FFalse
Question 4 True / False

Any image transformation that a human would still correctly label is a valid augmentation for model training — the primary requirement is that the label is preserved.

TTrue
FFalse
Question 5 Short Answer

Explain why domain knowledge is essential when choosing data augmentation strategies, using a specific example where an augmentation would be appropriate in one domain but harmful in another.

Think about your answer, then reveal below.