Which statement best describes data augmentation?

• A. It reduces model training time by discarding samples.
• B. It normalizes features to zero mean and unit variance.
• C. It increases dataset diversity by applying transformations to data samples.
• D. It directly optimizes the loss function without data changes.

Answer: (C)

Data augmentation expands the training data by applying plausible transformations to existing samples, creating new variations that preserve the label. This exposes the model to a wider range of appearances and conditions, helping it learn patterns more robustly and generalize better to unseen data. That’s why increasing dataset diversity through data transformations best describes the concept.

In practice, for images you might flip, rotate slightly, crop, or adjust brightness and color. For audio you could alter speed or pitch, and for text you might use careful paraphrasing or synonym replacement, all while ensuring the label stays the same. These transformations introduce variety without changing the underlying meaning or category.

The other ideas aren’t data augmentation: discarding samples to speed up training reduces the dataset rather than augmenting it; normalizing features to zero mean and unit variance is a preprocessing step, not creating new data; directly optimizing the loss without changing data focuses on training dynamics, not data variation.