Recently, many applications for the Restricted Boltzmann Machine (RBM) have been developed for a large variety of problems. However, RBMs are typically not considered as stand-alone solutions to classification problems, but are instead used as feature extractors for some other learning algorithm or to provide a good initialization for deep feed-forward neural network classifiers.

In this talk, I will argue that an RBM can directly address the problem of classification through its extension: the Classification Restricted Boltzmann Machine (ClassRBM). I will present different algorithms for training the ClassRBM, discuss the concepts of discriminative and generative learning and explain how they can be combined to obtain improved performances. I'll also cover the standard setting of multi-class classification as well as the semi-supervised and multi-task (or multi-label) settings. Using these algorithms, state of the art performances can be achieved by the ClassRBM on a variety of datasets, such as image, text document and music audio clip classification problems. Finally, I'll describe a further extension for vision tasks, in which simulated eye fixations can be combined to classify images. This approach allows the model to ignore irrelevant parts of an image and focus on its informative regions. An application to facial expression recognition will be presented.