Interpretable Analysis and Pruning of Modulation Recognition Network Based on Deep Learning

Abstract

Concerning poor interpretability and complexity of deep model in modulation recognition (MR) based on deep learning, an interpretable analysis and pruning framework of MR network based on Gradient-weighted Class Activation Mapping (Grad-CAM) is accordingly proposed in this paper. The framework first extracts the amplitude, phase and spectrum from the original modulated signal, and it uses the Smoothed Pseudo Wigner-Ville Distribution (SPWVD) to obtain the two-dimensional time-frequency spectrum of the modulated signal. Then, the key features in the deep model are visualized from the perspective of one-dimensional features and two-dimensional features at the input respectively. The framework visually displays and compares the differences and commonalities of the depth features of hidden layer with different models, extract the values of different filters of each layer in the deep neural network (DNN), and prune the network according to the values. The experiment results show that the interpretable and pruning framework of MR network based on Grad-CAM in this paper can achieve effective explanation and analysis on the MR network, and can greatly reduce the redundancy of the network. The running speed of the pruned network is 3.83 times higher than that of the original network. The size of the pruned network is 72% lower than that of the original network. Besides, the accuracy of the pruned network is 0.3% higher than that of the original network.