Multiple sparse spaces network pruning via a joint similarity criterion.

In this paper, a simple and effective neural network pruning framework is proposed to solve the problems of low model acceleration efficiency and inaccurate identification of pruning channels in conventional methods. Therefore, this paper first proposes a multi-sparse space network pruning scheme, which reduces the impact of pruning on network performance by defining the pruning task as an optimisation task in two different sparse spaces to gradually remove redundant parameters from the network. In this paper, we focus on the distribution characteristics of network weights in different sparse spaces, and we show that a decision method combining distance and direction information between weights can better locate the redundant information in the network. Experimental results and analysis have shown that the method can effectively prune neural networks, obtaining better results at higher compression and acceleration rates compared to other state-of-the-art methods. For example, on CIFAR-10, it reduces FLOPs by 67.5% and 64.2% for ResNet56 and ResNet110, respectively, while improving accuracy by 0.10% and 0.55%, respectively. On the CIFAR-100 dataset, the FLOPs for ResNet32 were reduced by 40.3%, while the accuracy was improved by 0.06%. On the STL-10 dataset, it was able to reduce the FLOPs of the ResNet18 model by 71.5% and gain an accuracy improvement of 0.59%. [ABSTRACT FROM AUTHOR]