Power Reduction in CNN Pooling Layers with a Preliminary Partial Computation Strategy

Abstract

Convolutional neural networks (CNNs) are responsible for many recent successes in the computer vision field and are now the dominant approach for image classification. However, CNN-based methods perform many convolution operations and have high power consumption which makes them difficult to deploy on mobile devices. In this paper, we propose a new method to reduce CNN power consumption by simplifying computations before max-pooling layers. The proposed method estimates the output of the max-pooling layer by approximating the preceding convolutional layer with a preliminary partial computation. Then, the method performs a complementary computation to generate an exact convolution output only for the selected feature. We also present an analysis of the approximation parameters. Simulation results show that the proposed method reduces the power consumption by 21% and the silicon area by 19% with negligible degradation in classification accuracy for the CIFAR−10 dataset.