Loss Functions#

Twitter Handle LinkedIn Profile GitHub Profile Tag

Learning is impossible without a way to measure how wrong a hypothesis is. A loss function assigns to every prediction a non-negative penalty, and the machinery of empirical risk minimization — and therefore of training itself — consists of minimizing its average over data.

This chapter first defines loss functions formally, including the view of a loss as a random variable, which is what licenses talking about the expected risk. It then studies two workhorses rooted in information theory: entropy together with information gain, and the cross-entropy loss.

Table of Contents#

References and Further Readings#