What data augmentations help small image datasets?

WHAT THIS TESTS The interviewer checks whether you understand augmentation as a regularizer that encodes invariances, and that the transforms must preserve the label and match the domain.

A GOOD ANSWER COVERS With little data a high-capacity network memorizes the training set and generalizes poorly. Augmentation synthetically expands the data by applying transformations that keep the class label valid, teaching the model invariances it should have. Common geometric transforms include random crops and resizing, horizontal flips, small rotations, translations, and scaling. Photometric transforms include brightness, contrast, saturation, and hue jitter, plus added noise or blur. More advanced techniques include Cutout or random erasing, Mixup which blends two images and their labels, CutMix, and automated policies like RandAugment or AutoAugment. The key principle is label preservation and domain appropriateness; you also apply augmentation only to the training set, never the validation or test set, so evaluation reflects real performance.

COMMON WRONG ANSWERS Using transforms that invalidate the label, like vertically flipping a digit so a 6 becomes a 9, or strong rotations on text. Augmenting the test set. Treating augmentation as a cure-all that replaces collecting more real data.

LIKELY FOLLOW-UPS Domain-specific choices, why Mixup helps calibration, test-time augmentation, and when augmentation hurts if it distorts the data distribution.

ONE CONCRETE EXAMPLE For a small cat-versus-dog set, random crops, horizontal flips, and mild color jitter give the network many varied views of each animal, so it learns that a dog is a dog regardless of position or lighting, cutting validation error without any new labeled photos, while a vertical flip would be left out since upside-down pets are unrealistic.