Approximated Prediction Strategy for Reducing Power Consumption of Convolutional Neural Network Processor

Takayuki Ujiie, Masayuki Hiromoto, Takashi Sato

Abstract: Convolutional neural network (CNN) is becoming popular because of its great ability for accurate image recognition. However, the computational cost is extremely high, which increases power consumption of embedded CV systems. This paper proposes an efficient computing method, LazyConvPool (LCP), and its hardware architecture to reduce power consumption of CNN-based image recognition. The LCP exploits redundancy of operations in CNN and only executes essential convolutions by an approximated prediction technique. We also propose Sign Connect, which is a low computational-cost approximated prediction without any multiplications. The experimental evaluation using image classification dataset shows that the proposed method reduces the power consumption by 17.8%-20.2% and energy consumption by 11.4%-14.1% while retaining recognition performance.