Deep learning methods for error correction

Abstract

Error correcting codes are used in communication systems to protect the transmitted information from noise and enable reliable information exchange between senders and receivers. Significant research effort has been made to devise well performing channel codes. However, traditional channel coding is not always optimal when the noise channel cannot be precisely modelled analytically. Deep Neural Network autoencoders have been used in place of traditional methods for both source and channel coding, as well as Joint Source-Channel Coding (JSCC). These deep learning methods are able to perform as well as or better than traditional methods, but also rely on an analytically modelled noise channel. In this work we develop a deep learning method for stochastic noise channel approximation using Generative Adversarial Networks (GAN) that can then be used for training of deep autoencoders for transmission over a complicated noise channel. We show that the proposed method is effective for a real-world noise channel, and successfully carry out JSCC autoencoding to transmit images over this channel.