Deep Boltzmann machines (DBMs) are exciting for a variety of reasons, principal among which is the fact that they are able to learn probabilistic representations of data in an entirely unsupervised manner. This allows DBMs to leverage large quantities of unlabelled data which are often available. The resulting representations can then be fine-tuned using limited labelled data or studied to obtain a more comprehensive understanding of the data at hand.

This talk will begin by providing a high level description of DBMs and the training algorithms involved in learning such models. A topic modelling example will be used as a motivating example to discuss practical aspects of fitting DBMs and potential pitfalls. The entire code for this project is written in python using only standard libraries (e.g., numpy).