Compression Algorithm for End-to-End Communication using CNN

Autoencoder has been applied in end-to-end physical layer communication, and its performance exceeds that of traditional communication systems in some scenarios. However, autoencoder has very high computational complexity and requires a lot of storage resources. This makes them difficult to deploy on embedded systems with limited hardware resources. In this paper, we introduce model compression algorithm for the end-to-end communication system composed of convolutional neural networks (CNNs). Firstly, the pruning algorithm for end-to-end communication system is designed to delete the weight coefficients that are not important to the performance. Secondly, according to the characteristics of autoencoder, we design a codebook-based quantization scheme to further reduce the memory consumption. In the case of Gaussian, Rayleigh and bursty fading channels, the experimental results show that the number of weight coefficients and the required resources are greatly reduced after compression, without performance degradation in low Eb/N0 and negligible performance degradation in high Eb/N0.