Suppose that you’re trying to build a hotdog classifier, and you have a choice between two neural networks: network A has a cross-entropy loss of 0.01 on the training set, and it reached that loss extremely quickly. Network B obtains a cross-entropy loss of 0.0001 on the training set, but it took a lot longer to get below 0.01 than Network A. You want to use the model that will be the most accurate on new, unseen possible-hotdogs: which do you pick?