Deep Neural Networks on Chip - A Survey

Currently, deep neural networks (DNNs) are widely used for various applications and have achieved state-of-the-art performances. A survey about the prior researches addressing efficient processing of DNNs in hardware domain is presented in this paper. Firstly, the concept and conventional structure of DNNs is introduced. Secondly different hardware platforms that used for DNNs on chip processing are emphasized in detail. Thirdly, it presents the main approaches for achieving energy efficiency and decreasing the computational cost, while maintaining accuracy for hardware implementation. Finally, the explorable research directions are discussed. From this work, this survey can provide a reference to the hardware researchers in the area of neural networks.