It's not the most useful outcome for the end user, but it's the perfect outcome from the perspective of the learning algorithm. All these things care about is m

It's not the most useful outcome for the end user, but it's the perfect outcome from the perspective of the learning algorithm. All these things care about is minimizing their loss function, where loss is deviation from the training set.