Modified Autoencoder Structure for Reducing BER Performance

Autoencoders have shown a great interest in digital communications in the last few years. Autoencoders based on deep neural network layers are able to generate the modulator at the emitter side together with the digital receiver matched to channel conditions. Furthermore, error correcting capabilities can be added to the autoencoder. As long as Symbol Error Rate (SER) is concerned, autoencoders outperform classical digital modulations as M-QAM. Nevertheless, because Gray coding is not embedded in autoencoders training, their Bit Error Rate (BER) is worse than their classical Gray coded M-QAM equivalent modulation. In this paper, we propose an improvement of classical autoencoders offering Gray coding capabilities and showing better BER than its classical M-QAM counterpart. In our work, we have modified the autoencoder model proposed in [1] by adding pre and post-processing layers in order to map constellation points on a Gray like map and we have modified the learning loss from categorical cross-entropy to binary cross-entropy. Simulation results show that the BER performances of this modified autoencoder are better than the ones of classical M-QAM systems. We also present the evolution of the autoencoder decision regions for different values of the training SNR that can explain their enhanced BER performance.