Circular Convolutional Auto-Encoder for Channel Coding

In this article, we investigate deep auto-encoders for channel coding to combat the curse of dimensionality commonly existing in learning based coding. Inspired by convolutional codes, we use hierarchical convolutional layers in both the encoder and the decoder. Codes with rate 1/2 and length 100 is learned automatically by the auto-encoder and have shown better performance than the convolutional codes. Our investigation indicates that the curse of dimensionality can be overcome by using the convolutional layers. We also show that the deep auto-encoder can learn to handle nonwhite noise automatically when the channel goes beyond additive white Gaussian noise (AWGN).