Functionally-Predefined Kernel: a Way to Reduce CNN Computation
Yuta INOUCHI, Hayato Yamaki, Shinobu Miwa, Tomoaki Tsumura · 2019
Convolutional Neural Networks (CNNs) have achieved high classification accuracy in image recognition, and now, they are widely used for numerous applications. For higher accuracy or more advanced applications, CNNs need to consume tremendous computational resources and time. Hence, many studies for reducing the computational cost of CNNs are actively being conducted. However, many previous methods for reducing the computational cost lead to a non-negligible loss in output accuracy. Therefore, it is still a challenge to reduce the computational cost of CNNs with keeping the output accuracy high. In this paper, we propose a novel concept "Functionally-Predefined Kernel" to reduce the computational cost for CNN training and discuss the potential of computation reuse to reduce the computational cost for CNN inference. Our experimental results show that the number of parameters to be trained can be significantly reduced by utilizing Functionally-Predefined Kernels without accuracy loss. In addition, we revealed that CNN's inference process includes many convolution operations with the same inputs and computation reuse, therefore, has high affinity to CNN computation.