Unsupervised PulseNet: Automated Pruning of Convolutional Neural Networks by K-Means Clustering

Convolutional Neural Networks (CNNs) achieve state-of-the-art results in many application areas, including image classification. For some applications it would be useful but impractical to deploy them on mobile devices with limited memory and power. A currently active area of research is the compression of deep networks while maintaining accuracy, with the aim of reducing memory usage, energy consumption and processing time. Several network compression methods have been proposed and have achieved good results, but they usually require the specification of parameters and are computationally expensive. We propose a new fast automated method called Unsupervised PulseNet that uses unsupervised k-means clustering to detect clusters of similar filters, and nodes in fully-connected layers, and prunes those that are redundant. We evaluate it on the CIFAR10, CIFAR100 and Tiny-Imagenet datasets using Alexnet, VGG16 and a 2-layer CNN called CifarNet suggested by the Tensorflow group. Compared to other methods in the literature we achieve the greatest compression, in shorter times, and with negligible loss in classification accuracy. In particular, we reduced Alexnet down to less than 0.7% of its original size, while not losing more than 2% classification accuracy.