Learning representations: a deep architecture based approach

Abstract

Representation Learning has become an active topic of research in the recent years. Neural models have been specially successful in this area due to recent developments in both algorithms and architectures. It is well known that by stacking layers of non linear functions we can learn increasingly complex representations which are able to capture the factors of variance in the data. In this work we develop a library containing feature learning algorithms using the Restricted Boltzmann Machine model as a building block. We test the developed algorithms with a simple dataset such as MNIST. Later we apply some single layer and multilayer feature learning algorithms in order to test whether we can learn useful representations which can be helpful when applied in a complex classification task such as Gender recognition. Finally, we explore a method to improve the classification results in a class imbalanced dataset such as the one used for the experiments.