A Comparison of Supervised and Semi-supervised Training Algorithms of Restricted Boltzmann Machines on Biological Data

Most devices applied nowadays in biological experiments produce large quantities of data. However, due to the cost of human expertise, only a small subset of these data can be examined or labeled by humans, while a huge part of the extracted information remains unlabeled. Therefore, interest in machine learning methods that can automatically make use of unlabeled data (sometimes in parallel with the annotated data parts) have grown steadily in the past years. In this study we show that one approach to handle the issue stated above could be the application of Restricted Boltzmann Machines (RBMs) in a semi-supervised machine learning configuration. RBMs can extract relevant information from the unlabeled data, and use it to improve the classification accuracy achieved by learning only on the labeled data set. We compare semi-supervised classification RBMs with a (deep) Artificial Neural Network and a Support Vector Machine to examine the improvements originating from the use of the unannotated data in three different, biologically inspired data sets. Our results show that even in this modern era of deep learning, semi-supervised classifiers are a viable option in these areas of science where annotating data is a time-consuming and resource-heavy task.