Skip to content
/ ibp Public

Take 3 on inference for Indian Buffet Processes

Notifications You must be signed in to change notification settings

rachtsingh/ibp

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Experimental research code

This repository attempts to implement several kinds of inference for generative models with an Indian Buffet Process prior (e.g. nonparametric VAEs, etc.). There's a strong focus on correct implementations and testing, since it is extremely easy to make a mistake in this area.

Inference methods implemented:

  • Gibbs sampling
    • collapsed Gibbs (i.e. marginalize out A) (Griffiths & Ghahramani (2005), though the derivation is in Doshi-Velez (2009))
    • uncollapsed Gibbs
  • Slice sampling (unimplemented)
  • Variational inference (VI)
    • Coordinate-ascent VI (CAVI) - the "older" method of VI (derived in Doshi-Velez (2009))
    • Stochastic VI (SVI) - see Hoffman (2013), essentially subsampling data
    • Autograd on the exact ELBO (ADVI-exact): from Doshi-Velez (2009) we have an exact way to compute the ELBO (without sampling from q), we can do essentially gradient-based maximum likelihood estimation here. This is extremely fast and stable (this method is implemented in Pyro as well)
    • Autograd on sampled ELBO - as a comparison, we sample some variables from q
    • Amortized VI (AVI or VAE) - we also fit a map from x to q(z; \lambda(x)), which helps scale obviously (number of parameters is fixed rather than O(n)) but also allows data sharing.

A very good summary of various methods Zhang - Advances in Variational Inference, 2017

Liu - Stein Variational Gradient Descent: A General Purpose Bayesian Inference Algorithm, 2016

Ranganath,Altosaar,Tran,Blei - Operator Variational Inference, 2016

Altosaar, Ranganath, Blei - Proximity Variational Inference, 2017

Vikram, Hoffman, Johnson - LORACs prior for VAE, 2018

About

Take 3 on inference for Indian Buffet Processes

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published