Skip to content

GSoC'20 with NumFOCUS

I am super excited to say that I have been selected as a Google Summer of Code student by NumFOCUS for PyMC4. I would like to thank my mentors Thomas Wiecki and Maxim Kochurov and the entire NumFOCUS community for giving this opportunity.

My project is about adding Variational Inference Interface to PyMC4. Variational Inference scales better over larger datasets as compared to the traditional MCMC algorithms. First, I had plans to implement OPVI1 framework as done in PyMC3 this summer. But as corrected by my mentor Maxim Kochurov, it would have taken extra time and more debugging because of the difficulty to deal with symbolic graph manipulations in Tensorflow. Now, the whole plan is to implement two Variational Inference Algorithms - Mean Field ADVI2 and Full Rank ADVI2 in PyMC4. Mean Field ADVI posits a Spherical Gaussian family and Full Rank ADVI posits a Multivariate Gaussian family to minimize KL divergence. I will write a blog post upto next week explaining both these algorithms.

All in all, I look forward to a great summer.

  1. Operator Variational Inference Rajesh Ranganath, Jaan Altosaar, Dustin Tran, David M. Blei (2016) ↩

  2. Automatic Differentiation Variational Inference Alp Kucukelbir, Dustin Tran, Rajesh Ranganath, Andrew Gelman, David M. Blei (2016). ↩↩