Mixup and CutMix Augmentation Quiz

Answers and explanations appear as you take the quiz.

1. How does Mixup create a new training example?
   • By rotating a single image and keeping its label
   • By cropping a random square and discarding it
   • By converting the image to grayscale
   • By linearly interpolating two images and their labels with a weight lambda
2. What is the core difference between CutMix and Mixup?
   • CutMix uses only one image while Mixup uses two
   • CutMix pastes a rectangular patch from one image onto another instead of blending pixels everywhere
   • CutMix changes only the label and never the image
   • CutMix applies exclusively to text data
3. In CutMix, how is the mixing weight (lambda) for the labels determined?
   • It is always fixed at 0.5
   • By the brightness of the brighter image
   • By the proportion of pixel area each image contributes
   • Randomly, with no relation to the images
4. Besides the input image, what else do Mixup and CutMix modify?
   • The label, turning it into a soft mixed target
   • The optimizer's learning rate
   • The number of layers in the network
   • The batch size
5. Which distribution is commonly used to sample the mixing coefficient lambda?
   • Uniform distribution over integers
   • Beta distribution
   • Poisson distribution
   • Bernoulli distribution with p fixed at 1
6. A documented benefit of these mixing augmentations is improved:
   • Compression of the trained model file
   • Training speed by removing the need for backpropagation
   • Ability to train without any labels
   • Robustness and calibration in addition to accuracy
7. Mixup uses lambda = 0.25 for image A and one minus lambda for image B. What target weights follow?
   • 25% A and 75% B
   • 75% A and 25% B
   • 50% each regardless of lambda
   • 100% A because it is listed first
8. A CutMix patch replaces 20% of image A’s pixel area with image B. What label weights does the stated area rule use?
   • 20% A and 80% B
   • 100% B for any patch
   • 80% A and 20% B
   • 50% each for all rectangles