Hardware Implementation Of Feedforward Artificial Neural Network Based On FPGA

Artificial neural network(ANN)is widely used in medicine,transportation,economy,information and other fields because of its outstanding ability of parallel distributed and adaptive information processing.Among them,feedforward artificial neural network is the most classical model of artificial neural network.This paper focuses on the realization of feedforward neural network with reusable neurons on FPGA.Firstly,This paper describes the research background,significance and research status at home and abroad.The main difficulties are as follows: There are many neurons in the artificial neural network model,and each neuron corresponds to multiple weight information.It needs a lot of hardware resources to realize all the neurons in artificial neural network and to store the corresponding weight information of neurons.Secondly,this paper studies and designs a feedforward artificial neural network with reusable neurons.On this basis,the model of feedforward artificial neural network is compressed by combining with the principle of pruning algorithm.In this paper,the time-sharing multiplexing of neuron group module is used to realize the transmission of information among different layers,there is no need to realize all the neurons in the artificial neural network,so as to reduce the hardware resource consumption.According to the multiplication and accumulation operation in the linear operation of artificial neural network feedforward algorithm,overflow protected adder and overflow protected base 4 Booth multiplier are designed.The base 4 Booth algorithm can reduce the consumption of storage resources by reducing partial product,at the same time,the fixed number is used to improve the operation efficiency.In this paper,a look-up table method is used to realize the nonlinear Sigmoid function.All the neurons in the neuron group reuse a LUT module in time-sharing,without the need to design a LUT module for each one,so as to effectively reduce the cost of hardware resources.Finally,the FPGA verification platform is built.The RTL level simulation verification is completed by using ISE development kit,and the synthesis and timing analysis of the whole feedforward artificial neural network circuit are completed on Vivado development kit.Then the prototype verification is completed based on Xilinx artix-7 FPGA development board.We have realized that the feedforward artificial neural network model composed of78 neurons need 16066 LUTS and 18936 FF.At the clock frequency of 200 MHz,it takes26.5 ?s time to complete the information transfer from input layer to output layer in feedforward artificial neural network,and the overall power consumption of the circuit is341 m W.