Hierarchies of probability vectors can be modelled, and learning/inference done using the Dirichlet Process and the Pitman-Yor Process. Discrete feature vectors (Booleans, counts, etc.) can be modelled using non-parametric versions of vectors of Bernoulli-Betas, or Poisson-Gammas, etc. (including the well published Indian Buffet process).
This tutorial will present these. The first part will be done informally on the board from notes and will give a machine learning perspective on the statisticians' models for discrete feature vectors. This will present Lancelot James' theory (http://arxiv.org/pdf/1411.2936) for easier consumption. The second part will present some of our theory of learning on hierarchies of probability variables from the longer tutorial
http://topicmodels.org/2014/10/30/a-tutorial-on-non-parametric-methods-for-probability-vectors
Many authors are now proposing the former (non-parametric models of discrete feature vectors)
be used as a component for matrix models like topic models.
About the speaker:
http://infotech.monash.edu.au/research/profiles/profile.html?sid=6245956&pid=10352
Last updated on 12 Jan 2015 by Antti Ukkonen - Page created on 12 Jan 2015 by Antti Ukkonen