From the Deep Learning Specialization, I learned that dev and test set should be similar. Andrew gave the analogy that the dev set is the target we aim at using the train set, and the test set is the actual target we want to hit. If that is the logic, why would we augment validation data? With or without test data, we want the validation data to be close to the distribution of the true data, so that we are aiming at the right spot with the train set. Wouldn’t augmenting validation distort it from the true data distribution?
We shouldn’t augment validation data. The only processing that’s allowed for validation data is to divide each pixel by 255.