Random binary vectors
The class of distributions that cause you to reinvent Shannon information if you stare at them long enough
2017-02-20 — 2022-03-30
Wherein distributions over n-length binary vectors are examined, and the existence of 2^n possible outcomes is noted, with continuous Gumbel-softmax relaxations and piano-roll representations being described.
Distributions over random boolean vectors. Useful in computer science and piano rolls. Not quite the same as categorical distributions, although those can be written as distributions over boolean vectors. In a multi-class classification case, each realisation has only one class; in an \(n\)-class rv, there are \(n\) possible outcomes. In a multivariate Bernoulli distribution, there are \(2^n\) possible outcomes.
1 Continuous relaxations
Multivariate Gumbel-softmax tricks.
2 Paintbox models
Not sure how these work but maybe related. See (Broderick, Pitman, and Jordan 2013; Zhang and Paisley 2019).
3 Matrix models
TBC.
See, e.g. Lumbreras, Filstroff, and Févotte (2020)