6. Components
1. Augmentations
• Without corruptions in test data
• Severity (eg. rotate: 2°, -15°)
• Length of chain: 1~3 (uniformly random)
2. Mixing
• Elementwise convex combinations
• Sample mixed images coefficients from Dirichlet(a, …, a)
• Sample skip-connection weight from Beta(a,a)
3. Jensen-Shannon Divergence Consistency Loss
• Enforces smoother neural network responses
• Stochasticity
• the choice of operations, the severity of these operations, the lengths of the
augmentation chains, and the mixing weights
8. Objective
• Dataset:
• Training: Cifar100
• Test: Cifar100-C (#corruption =n, n images + 1 label)
• Base Model: WideResNet 40-2
• Task:
1. Compare average classification error between
• AugMix w/ JSD loss of 1/2/3 augmixed image(s)
• AugMix w/o JSD loss
• No AugMix
2. Show some augmented images
3. (optional task) What if change black blank area to
white or random pixel?