Compression of Deep Convolutional Neural Networks for Fast and Low Power Mobile Applications

Although the latest high-end smartphone has powerful CPU and GPU, running convolutional neural networks (CNNs) for complex tasks such as ImageNet on mobile devices is challenging. To deploy deep CNNs on mobile, we present a simple and effective scheme to compress the entire CNN, we call one-shot whole network compression. The proposed scheme consists three steps: (1) rank selection with variational Bayesian matrix, (2) Tucker decomposition on kernel tensor, and (3) fine-tuning recover accumulated loss of accuracy, and each step can be easily using publicly available tools. We demonstrate the effectiveness of proposed scheme by testing the performance of various compressed CNNs (AlexNet, VGGS, GoogLeNet, and VGG-16) on the smartphone. Significant in model size, runtime, and energy consumption are obtained, at the of small loss in accuracy. In addition, we address the important level issue on 1? 1 convolution, which is a key operation of module of GoogLeNet as well as CNNs compressed by our proposed.