GitHub - yorkerlin/VB-MixEF: Code for ICML 2019 paper on "Fast and Simple Natural-Gradient Variational Inference with Mixture of Exponential-family Approximations"

yorkerlin / VB-MixEF Public

Notifications You must be signed in to change notification settings
Fork 6
Star 17

Code for ICML 2019 paper on "Fast and Simple Natural-Gradient Variational Inference with Mixture of Exponential-family Approximations"

17 stars 6 forks Branches Tags Activity

Star

Notifications

Branches Tags

Name		Name	Last commit message	Last commit date
Latest commit History 67 Commits
src/matlab		src/matlab
README.md		README.md
poster_main.pdf		poster_main.pdf
poster_workshop.pdf		poster_workshop.pdf
report.pdf		report.pdf

Repository files navigation

For an efficient mixture of Gaussians (MOG) implementaiton using vectorization（which is 20 times faster than this repo), please use this new repo based on our ICML 2020 paper.

Code for ICML 2019 paper on Fast and Simple Natural-Gradient Variational Inference with Mixture of Exponential-family Approximations

To-do List:
- added a poster and a technical report about the gradient identities used in the paper (To appear at the ICML workshop on Stein's method) [done]
- added a poster of the main paper. [done]
- [skewness] added a Matlab implementation for the toy example using skew Gaussian and exponentially modified Gaussian. [done]
- [multi-modality] added a Matlab implementation for the toy example using MoG. The implementation is based on this repo [done]
- [heavy tails] To add a Matlab implementation for BLR using t-distribution and symmetric normal inverse Gaussian
- To add a Python implementation for Vadam extensions