Filter Pruning in Convolutional Neural Networks Using Structural Similarity Based K-Means

Abstract

Convolutional Neural Networks (CNNs) have been recently seeing great success in various image classification fields and applications. However, this success has been accompanied by a significant increase in memory and computational demands, limiting their use in resource-limited devices, e.g., smartphones. In response, network pruning methods, in particular filter pruning, are seeing increased interest. The principal goal of the current pruning algorithms is to substantially reduce the resource demands for executing the forward pass of a trained CNN, while minimizing performance degradation. In this thesis, we propose a new approach for filter pruning in CNNs. Our filter pruning method utilizes K-Means clustering based on the Structural Similarity Index Measurement to group similar filters together in each convolutional layer. A representative filter is selected from each cluster and the remaining filters are considered redundant and pruned from the CNN. We evaluated our filter pruning method on the VGG-16 architecture with the benchmark CIFAR-10 dataset. We were able to reduce the computational demands (floating-point operations) of VGG-16 by over 50%. Simultaneously, the network’s performance remained significantly better than the one pruned by the HRank algorithm. The results of our experiments provide promising indications that our method can significantly outperform state-of-the-art filter pruning methods.