Filter pruning with uniqueness mechanism in the frequency domain for efficient neural networks

Filter pruning has drawn extensive attention due to its advantage in reducing computational costs and memory requirements of deep convolutional neural networks. However, most existing methods only prune filters based on their intrinsic properties or spatial feature maps, ignoring the correlation between filters. In this paper, we suggest the correlation is valuable and consider it from a novel view: the fre-quency domain. Specifically, we first transfer features to the frequency domain by Discrete Cosine Transform (DCT). Then, for each feature map, we compute a uniqueness score, which measures its prob-ability of being replaced by others. This way allows to prune the filters corresponding to the low -uniqueness maps without significant performance degradation. Compared to the methods focusing on intrinsic properties, our proposed method introduces a more comprehensive criterion to prune filters, further improving the network compactness while preserving good performance. In addition, our method is more robust against noise than the spatial ones since the critical clues for pruning are more concen-trated after DCT. Experimental results demonstrate the superiority of our method. To be specific, our method outperforms the baseline ResNet-56 by 0.38% on CIFAR-10 while reducing the floating-point operations (FLOPs) by 47.4%. In addition, a consistent improvement can be observed when pruning the baseline ResNet-110: 0.23% performance increase and up to 71% FLOPs drop. Finally, on ImageNet, our method reduces the FLOPs of the baseline ResNet-50 by 48.7% with only 0.32% accuracy loss.(c) 2023 Published by Elsevier B.V.