[2103.05909] A variational inference framework for inverse issues

View a PDF of the paper titled A variational inference framework for inverse issues, by Luca Maestrini and 1 different authors

View PDF
HTML (experimental)

Summary:A framework is offered for becoming inverse downside fashions through variational Bayes approximations. This system ensures flexibility to statistical mannequin specification for a broad vary of functions, good accuracy and diminished mannequin becoming occasions. The message passing and issue graph fragment strategy to variational Bayes that can be described facilitates streamlined implementation of approximate inference algorithms and permits for supple inclusion of quite a few response distributions and penalizations into the inverse downside mannequin. Fashions for one- and two-dimensional response variables are examined and an infrastructure is laid down the place environment friendly algorithm updates primarily based on nullifying weak interactions between variables may also be derived for inverse issues in larger dimensions. A picture processing software and a simulation train motivated by biomedical issues reveal the computational benefit provided by environment friendly implementation of variational Bayes over Markov chain Monte Carlo.