Uso de Deep Learning para la codificación y decodificación en canales de una sola vía

Abstract

This work evaluates the application of deep learning techniques in the discovery of codes for one-way channels, specifically in a BSC. The main objective is to create codes comparable to conventional coding techniques, addressing the challenges posed by the inclusion of machine learning in the process. This document provides an exhaustive review of related work and the theoretical framework necessary to understand the problem, covering neural networks and conventional codes in binary symmetric channels. Then, experiments with deep neural networks are proposed to create encoder and decoder models for the channel, describing the architectures, coding rates, and key technical aspects. Linear layers, GRU, and LSTM networks are employed. The experimental results derived from each experiment graphically show the performance of the codes discovered by the neural models compared to conventional error-correcting codes in terms of BER measurement for specific values of q (the bit error probability of the channel). Finally, the inherent limitations of the binary symmetric channel and their impact on the development and results of this work are discussed. Additionally, future research directions that could help improve the results are outlined, including an interpretation of the discovered codes within the framework of Coding Theory.