Norm-Correlation Based Filter Pruning to Accelerating Networks

recently, high-performance deep learning models have been developed in various fields. Model compression research is also underway to apply deep learning models to edge device such as mobile phone and embedded platforms. Representatively, pruning is a popular model compression method to accelerate the inference process of deep neural network. So we accept channel pruning based method to reduce memory consumption and model complexity. Our method considers norm importance of filters and the correlation between filters, and select filter that has small norm with large correlation. This approach reduces redundancy of filter and preserves only important filters. Norm-Correlation based pruning not only preserves uncorrelated filters that extract unique features with small importance, but also removes replaceable filters with large importance. Our method is applicable to Convolutional Neural Networks (CNNs) and on ILSCRC-2012, we reduce 30% parameters and 41.1% FLOPs of ResNet-34 within 1% top-5 accuracy loss.