MSc Informatics Dissertation - University of Edinburgh

This repo contains the thesis and code for my MSc dissertation, done through the School of Informatics at the University of Edinburgh. My supervisor was Dr. Michael Gutmann.

Thesis

The title of my thesis is Fast and Scalable Factor Analysis Algorithms for Bayesian Deep Learning.

It can be found here.

Algorithms

Online SGA FA

Algorithm 1 from the thesis is implemented in the OnlineGradientFactorAnalysis class here.

Online EM FA

Algorithm 2 from the thesis is implemented in the OnlineEMFactorAnalysis class here.

VIFA

Algorithm 3 from the thesis is implemented in the FactorAnalysisVariationalInferenceCallback class here.

Experiments

The main experiments from the thesis are defined in the dvc.yaml file.

After installing the dependencies in requirements.txt, the experiments can be reproduced by running the following command from root of the project:

dvc repro

In order to reproduce a specific experiment, run the following command:

dvc repro <stage name from dvc.yaml>

The stages correspond to the experiments in the following section of the thesis:

• online_fa an online_fa_analysis: Section 4.4.1
• linear_regression_vi: Section 5.2.2
• neural_net_predictions: Section 5.2.3

If you want to change any of the parameters for the experiments, you can do so in params.yaml.