Graph Structure Learning-Based Compression Method for Convolutional Neural Networks

Convolutional neural networks (CNNs) have achieved remarkable performance in diverse applications. Nevertheless, the substantial scale and computational intricacy limit the practical implementation of CNNs, particularly on resource- starved devices. This paper presents a compression technique based on graph structure learning (GSL) for CNNs. This method aims to capture the correlations among parameters in each neural network layer and compress the model by leveraging the strength of these correlations. Firstly, the neural network parameters are modeled as a graph structure by adopting a graph construction methodology. Subsequently, the graph is fed into a dual-branch GSL module. This module introduces a constraint that optimize and refine the original graph topology and maximize the difference in feature information obtained between the two channels. Through the process of graph learning, the existing correlations among the parameters of the CNNs are demonstrated. Finally, based on the correlations between the parameters of the CNNs, the parameters with lower relative importance are selected and the neural network parameters are compressed. The proposed method significantly reduces the parameter and floating-point computation complexity of CNNs, thereby diminishing the model’s intricacy. Furthermore, the operational efficiency of the network model is improved without compromising its prediction accuracy. The effectiveness of the proposed method is validated on the VGG-16 and ResNet-101 models. The compressed models’ accuracy, efficiency, and memory consumption are then compared with the original models to demonstrate the effectiveness of the method.