Mejoras en tratamiento de problemas de clasificación con modelos basados en autoencoders

Resumo

In summary, the main contributions of this thesis are as follows: A theoretical analysis and taxonomy of the main autoencoder variants present in the literature, composing a guide to ease their selection and use. A complete software package which automatizes a great part of the implementation work for autoencoders and simplifies its use to a level similar to other feature extraction methods. A synthesis and organization work of the peculiarities that supervised learning problems can present when data points are represented in a nonstandard fashion. A demonstration of the diverse applications of autoencoder-based models, identifying and exposing several strategies to solve unsupervised problems by means of variable transformations. Three new models, Scorer, Skaler and Slicer, focused on data complexity reduction in classification problems. This document introduces all global concepts needed to understand the published articles and provides a theoretical vision of the representation learning problem and of the deep learning tool set, which includes the main object of study. In addition, it explains the techniques that help put into practice these models and how they execute on computation infrastructures. Next, the material published throughout the doctoral period is introduced and five articles published in renowned journals are reproduced. Finally, these and other activities carried out are summarized and the lines of work that would continue the achieved advancements are presented.