Introduction
A handful of named distributions cover most discrete data in ML: yes/no outcomes, counts of successes, choices among classes, and counts of events. Each one is also the output layer of a familiar model — sigmoid for Bernoulli, softmax for Categorical.