Convolutional Neural Network-Based Polar Decoding

In this work, we extend the capability of convolutional neural network (CNN) to polar code decoding. Previous work has shown that a multi-layer perceptron (MLP), which is a basic form of deep neural network (DNN), can achieve high decoding accuracy and speed for polar code when the block length is very short. However, its performance drops drastically for longer codes, due to the bulky network structure. In this work, we design and implement a CNN for polar decoding. In order to improve the decoding accuracy, we introduce auxiliary labels into CNN output based on the encoding structure of polar code. In addition, we propose to prune the CNN to preserve the decoding accuracy of wider network while reducing the computation and the parameters. With these two innovations, the decoding accuracy of original CNN can be improved. We carry out extensive simulations to compare our designed CNN decoder with MLP decoder. Results show that when the code length is 64, our model is 60 times smaller than the MLP decoder, and the accuracy of our model increases with model size, while MLP reaches saturation. Additional results show that our proposed method outperforms original CNN with regard to the BER performance under marginal parameter increase.