Activation Distribution-based Layer-wise Quantization for Convolutional Neural Networks

As convolutional neural network (CNN) research accelerates with advances in GPU s, the accuracy of CNN models has been continuously improved. However, in proportion to the enhancement of model accuracy, the computational amount of the CNN models increases, which causes a problem that it is difficult to practically use the CNN models in mobile/embedded platforms. To address this problem, optimization and weight reduction methods for CNN models have been actively studied. This paper proposes a new scale factor for layer-specific quantization considering the activation distribution of CNN s. The proposed method has the advantage that it is possible to minimize the accuracy drop for each layer and is friendly to hardware accelerator design. As a result, the proposed quantization method achieves much higher accuracy compared to the quantization studies on the conventional accelerator design while maintaining low hardware resources.