abegehr/ML_compare_mixture_algs

Comparing expectation maximization and variational inference algorithms for fitting Gaussian mixtures

ML_compare_mixture_algs

Goal: Compare expectation maximization (EM) and variational inference for mixture of Gaussians.

View animated example

Using the settings: n_components = 5, weight_concentration_prior = 1e-3, creates this animation comparing EM and variational inference for mixture of Gaussians:

Check out the whole Jupyter Notebook here: https://abegehr.github.io/ML_compare_mixture_algs/

… and then run your own experiments!

Run your own experiments:

1. Clone or download this repository.
2. Navigate to folder: cd (path to)/ML_compare_mixture_algs/
3. Start jupyter notebook: jupyter notebook
4. Webbrowser with jupyter notebooks running opens.
5. Open main.ipynb in jupyter notebooks.
6. Run all cells.
7. Animations comparing expectation maximization (EM) and variational inference applied on a mixture of Gaussians are shown. Great!
8. Change parameters to test the two methods. Use the settings in the second to last cell to change parameter settings under # PARAMETER SETTINGS HERE. For example, try these settings:
   • n_components = 4, weight_concentration_prior = 1e-3
   • n_components = 5, weight_concentration_prior = 1e-3
   • n_components = 8, weight_concentration_prior = 1e-3
   • n_components = 5, weight_concentration_prior = 1e+3
9. It is possible to experiment with many more parameters. Find the sklearn documentation for more information on possible parameters here:
   • Expectation Maximization (EM) for mixture of Gaussians: sklearn.mixture.GaussianMixture
   • Variational Inference for mixture of Gaussians: sklearn.mixture.BayesianGaussianMixture
10. Happy experimenting!

---

This animation was created as part of the seminar Mathematics of machine learning SS2018 (PD Dr. Pavel Gurevich). Group #6.

If you have any comments or questions, please contact: a.begehr@fu-berlin.de