The images in the CIFAR10 dataset are significantly more complex than the MNIST dataset. They are larger, have three times the amount of pixels and are more cluttered. This makes them harder to learn and consequently means we should use less augmentation.

less --> more?