Data Augmentation Quiz

Answers and explanations appear as you take the quiz.

1. What is the primary purpose of data augmentation?
   • To expand and diversify training data so models generalize better
   • To speed up inference on a trained model
   • To compress the size of the dataset on disk
   • To label unlabeled examples automatically
2. Which of these is a common image augmentation technique?
   • Gradient descent
   • Random horizontal flipping
   • Tokenization
   • Dropout
3. What does back-translation do in text augmentation?
   • Reverses the order of words in a sentence
   • Removes all stop words from the text
   • Translates text to another language and back to create paraphrases
   • Converts text into numerical embeddings
4. Why must augmentations be 'label-preserving'?
   • To keep the dataset file size constant
   • To make training run faster
   • Because models cannot read labels otherwise
   • Because changing the label would teach the model wrong associations
5. What does the Mixup technique do?
   • Blends two examples and their labels into one training sample
   • Masks a rectangular region of an image
   • Searches for the best augmentation policy automatically
   • Adds random noise to model weights
6. When is data augmentation especially valuable?
   • When you already have unlimited labeled data
   • When labeled data is scarce and collecting more is hard
   • Only when the model is fully trained
   • Only for tabular spreadsheet data
7. What does CutMix do to an image and its training label?
   • It adds a random word to a sentence without changing the label
   • It swaps only class names in the dataset
   • It replaces a patch with a patch from another image and mixes labels by the replaced area
   • It blurs every input and keeps a single label unchanged
8. How does AutoAugment select augmentation policies in its original paper?
   • It chooses transformations from a fixed human-written list without evaluation
   • It uses test labels to relabel training images
   • It selects operations from inference-time confidence only
   • It searches for policies using validation performance as a reward signal