The Shape of Being Wrong

You can build the most elegant neural network architecture in the world, and it will learn absolutely nothing until you answer one question first: wrong compared to what, exactly? That's what a loss f