Making your neural network output confidence scores that actually mean something without much effort.