Research On Accelerating Optimization And Automatic Generation Technology For Artificial Neural Network

In recent years,the remarkable achievements that depth learning has realized in computer vision and speech recognition field make Artificial Neural Network(ANN)algorithm more and more attractive.However,with the development of ANN model,the computational complexity becomes so serious that large-scale computing platform must be used for training and testing.Training speed and power consumption has been the main bottleneck limiting ANN algorithm applied in real life.So it is more attractive to implement the ANN accelerator on Field-Programmable Gate Array.This paper focus on the accelerating optimization and automatic generation technology for ANN.First of all,by analyzing the algorithm structure in detail,we extracted the calculation core of ANN algorithm.Secondly,we evaluated fixed-point representation of Extreme Learning Machine and AlexNet algorithm,the results show the availability of fixed-point format for ANN and can be of great benefit to decrease the resource,we created the simulation and analysis environment for any ANN algorithm.Then,It is more important that we designed the corresponding accelerator structure for these computation cores combining with the characteristic of the FPGA platform,firstly,we improved the routing performance by changing the storage of traditional convolution accelerator,secondly,we realized the better large-scale matrix multiplication accelerator by balancing the storage and resource and optimizing the computation,thirdly,we improve the S igmoid accelerator by changing the multiplication into shift operator when adopting the Piecewise Linear Approximation(PLA)method.Finally,we improved the adaptability of accelerator by further designing the automatic generator for generating accelerator based on all works,meanwhile,we provided friendly user interface and parameters searching algorithm.