Key Issues of FPGA Implementation of Neural Networks

Recent years, the artificial neural networks were noticeable on development. As the bridge between the theoretical research and application research, the hardware implementation technologies have developed rapidly, particularly in configurable FPGA implementation technologies. However we have found the shortcomings of the existing methods which need to be improved. This paper has taken research on the key issues of the FPGA implementation of neural networks, discussing on the following issues: data representation, inner-products computation, and implementation of activation function, storage and update of weights, nature of learning algorithm and design constraints. It also introduces some relatively mature and new methods and pointed out their deficiencies and future works.