Mixtures of multivariate Bernoulli distributions can be used to model binary random vectors.

Training in terms of maximum likelihood estimation is often performed by an expectation-maximisation algorithm. In this research we consider a Bayesian hierarchical model, estimated using a Gibbs sampler. Empirical results for the Bayesian approach are compared with a Bernoulli mixture model trained via expectation-maximisation. These models are evaluated by comparing area under receiver operating characteristic metrics on a handwritten-digit estimation task. Algorithms are derived so that models can be meaningfully compared.

All code used for the research was written for Python 3.6, with the following dependencies:

• numpy == 1.16.2
• matplotlib == 2.2.2
• scipy == 1.2.1

The implementation for the Gibbs and EM models are in gibbs.py, and the implementation for AUROC metrics and label switching can be found in metrics.py. The remaining code was the code written to calculate estimates and plot graphs for the report.