Decoding of Polar Code by Using Deep Feed-Forward Neural Networks

With the success of image classification problems, deep learning is expanding its application areas. In this paper, we apply deep learning to decode a polar code. As an initial step for memoryless additive Gaussian noise channel, we consider a deep feed-forward neural network and investigate its decoding performances with respect to numerous configurations: the number of hidden layers, the number of nodes for each layer, and activation functions. Generally, the higher complex network yields a better performance. Comparing the performances of regular list decoding, we provide a guideline for the configuration parameters. Although the training of deep learning may require high computational complexity, it should be noted that the field application of trained networks can be accomplished at a low level complexity. Considering the level of performance and complexity, we believe that deep learning is a competitive decoding tool.