Project for the course Bayesian Methods for Machine Learning at Skoltech

• Hamidreza Behjoo
• Jaspers W. Huanay

We study here a variational objective to simultaneously trade off effects of mass-covering,spread and outlier robustness. This is done by developing a variational inference objectiveusing the alpha-beta (AB) divergence, a family of divergence governed by two parametersand covering many of the divergences already used for VI as special cases.

Alpha-Beta divergence is investigated and compared with respect to other wellknown divergences in literature.

For details of the derivation see bml2020_report.pdf

To get the results showed above, run:

• bayesian linear regression.ipynb, for data wihtout modification
• bayesian linear regression_outlier.ipynb, for data wiht outliers added

• Regli, Jean-Baptiste, and Ricardo Silva. "Alpha-beta divergence for variational inference." arXiv preprint arXiv:1805.01045 (2018).