Empirical Evaluation of Variational Autoencoders for Data Augmentation

Abstract

Since the beginning of Neural Networks, different mechanisms have been required to provide a sufficient number of examples to avoid overfitting. Data augmentation, the most common one, is focused on the generation of new instances performing different distortions in the real samples. Usually, these transformations are problem-dependent, and they result in a synthetic set of, likely, unseen examples. In this work, we have studied a generative model, based on the paradigm of encoder-decoder, that works directly in the data space, that is, with images. This model encodes the input in a latent space where different transformations will be applied. After completing this, we can reconstruct the latent vectors to get new samples. We have analysed various procedures according to the distortions that we could carry out, as well as the effectiveness of this process to improve the accuracy of different classification systems. To do this, we could use both the latent space and the original space after reconstructing the altered version of these vectors. Our results have shown that using this pipeline (encoding-altering-decoding) helps the generalisation of the classifiers that have been selected.