On the Efficient Design of Neural Networks in Communication Systems

Various types of neural networks (NNs) have shown promising performance in communication systems. However, the low-latency implementation of these tasks is currently impractical due to the high computational complexity and large model size of NNs. In this paper, we propose an iterative optimization framework with retraining process to adaptively find the quantization scheme for different NNs. Moreover, the efficient design of convolutional neural networks is presented to reduce the required parameters and computational complexity. Experiment results for modulation classification, channel decoder and equalizer are presented. Compared to the original full-precision models, the quantized NN models achieve comparable performance with only 4 to 5 weight bits and 8-bit activation. The size of optimized models is significantly compressed and the hardware complexity of the NN inference is also reduced.