These notes adapted from:

• "The Indian Buffet Process: An Introduction and Review", Griffiths, Ghahramani, 2011
• "Introduction to the Dirichlet Distribution and Related Processes", Frigyik, Kapila, Gupta
• "Advanced Data Analysis from an Elementary Point of View", Cosma Rohilla Shalizi, chapter 19, "Mixture Models"
• "Mixture Models and the EM Algorithm", slide presentation by Bishop 2006
• "Accelerated Sampling for the Indian Buffet Process", Doshi-Velez and Ghahramani

The tutorial by Griffiths and Gahramani above was my primary resouce. Then, in order to understand it, I needed to reach out to the other resources above :-)

Generally speaking, these notes assume that you are reading the appropriate tutorial/paper/slides in parallel with these notes. How this is organized currently:

The following sections go through Griffiths and Ghahramani tutorial, in sequence. Occasionally, I found I needed some additional resources/information, to understand the Griffiths and Ghahramani tutorial, and interlude to such other resources, in the middle of some of these sections.

• ibp_section1.ipynb revision of multinomial distributions, Dirichlet distributions, mixture models
• ibp_section2.ipynb Latent class models
• ibp_section3.ipynb Latent feature models
• ibp_section4.ipynb Example: Linear-Gaussian Latent Feature Model with Binary Features (in progress)

Some notebooks are apart from the above sections, since they use a ton of browser memory, so I've separated them out:

• [sampling from prior.ipynb](sampling from prior.ipynb) notebook trying some simple sampling from an IBP prior, without any data
• [Demonstration on 6x6 images.ipynb](Demonstration on 6x6 images.ipynb) notebook trying to sample from the posterior, with some simple toy data (from the Griffiths and Ghahramani tutorial). This is in progress for now

I also separated out my reading through Doshi-Velez's "accelerated sampling" paper, because it was also using a lot of browser memory. Actually, this will probably be moved to a separate repo at some point, probably.

• accelerated_gibbs_sampling.ipynb (This is in-progress, for now)