Which technique increases training data diversity through small automatic transformations?

Which technique increases training data diversity through small automatic transformations?

• A. Data augmentation
• B. Data acquisition
• C. Data cleansing and transformation
• D. Feature engineering

Answer: (A)

Data augmentation is the technique that increases training data diversity by applying small automatic transformations to existing data. By creating varied but labeled versions of the same samples, it helps the model learn to generalize to new, unseen data without needing to collect and label more data. In practice, image data augmentation might include rotations, flips, crops, or slight color changes; for text, lightweight paraphrasing or synonym replacements; for audio, tiny changes in speed or pitch. The key is that these transformations are label-preserving, so they expand the dataset while keeping the ground truth intact, reducing overfitting and improving robustness. Data acquisition focuses on gathering more real data, data cleansing and transformation centers on cleaning and normalizing data, and feature engineering builds new input features from existing data rather than increasing the number of labeled samples.

Data augmentation is the technique that increases training data diversity by applying small automatic transformations to existing data. By creating varied but labeled versions of the same samples, it helps the model learn to generalize to new, unseen data without needing to collect and label more data. In practice, image data augmentation might include rotations, flips, crops, or slight color changes; for text, lightweight paraphrasing or synonym replacements; for audio, tiny changes in speed or pitch. The key is that these transformations are label-preserving, so they expand the dataset while keeping the ground truth intact, reducing overfitting and improving robustness. Data acquisition focuses on gathering more real data, data cleansing and transformation centers on cleaning and normalizing data, and feature engineering builds new input features from existing data rather than increasing the number of labeled samples.