Reducing MAC operation in convolutional neural network with sign prediction

Due to recent researches on artificial neural net- work algorithms and machine learning, the accuracy of image recognition and natural language processing has increased to the level of human beings in specific fields. Especially, researches to improve the accuracy of algorithms are being actively conducted, and researches on hardware accelerators that implement such algorithms quickly and efficiently are actively under way. In order to utilize artificial intelligence reasoning ability as well as computing speed in mobile or embedded environment, it is necessary to reduce the power consumption and memory usage of artificial intelligence hardware. In this paper, we propose a algorithm to reduce the computational complexity in designing the CNN accelerator. We tried to reduce the MAC computation by encoding the inputs and predicting the sign of the MAC oper- ation. We confirmed the performance improvement by evaluating the sign predictor through the simulation results.