Loss-Functions

The Loss Function Is a Distribution Assumption

Choosing a loss is choosing a distribution for your output. Why every supervised network is a maximum-likelihood estimator, and why the gradient is always the residual.