The Research Of Neural Network Processor Based On STDP

The research of Artificial neural network technology has always been based on the biological nervous system,along with the continuous development and progress of biological neuroscience,people also try to improve the artificial neural network technology.With the development of technology,researchers have found that biological brains had lower frequency,lower power consumption and stronger robustness compared with computers,it also has smaller size,higher degree of parallelism,and better real-time performance.Based on observation and understanding of biological neural networks,researchers proposed Spiking Neural Network(SNN).Called third-generation neural networks by artificial neural network researchers,it is an artificial neural network that uses discrete neural pulse signals to process information and has high degree of real-time performance as well as good energy efficiency ratio.With the rapid development,it has attracted widespread attention from all over the world since it was proposed.On traditional computers,Spiking Neural Networks are implemented by software programming methods.However,Spiking Neural Networks is ought to do real-time calculations under many circumstances,which is not suitable by using software programming.Moreover,with smart portable devices and IoT products being rapidly developed,processors with lower power consumption and smaller volumes are urgent needed.For above reasons,the performance of Pulse Neural network algorithm in traditional computer can no longer meet the human needs.There is an urgent need for a platform that can validate and process SNN algorithms.The paper proposed a configurable SNN processor platform on the basis of STDP,which is designed with FPGA.Due to its universality and expansibility as well as the configurable internal structure and parameters,this platform can be applied to verifying and simulating for most SNN and SCNN algorithms,it can also save a lot of hardware resources with reusable design method.What's more,the paper applied the typical SNN algorithm for engineering-improved of the hardware platform,and tested the performance of the improved algorithm of the processor platform.The typical SNN algorithm is engineering-improved to be more suitable for the hardware-platform,and the improved algorithm is used to test the performance on the platform.Firstly,we introduced the background and significance of this study,and explained the structure and content of the paper.Secondly,we introduced the structure and characteristics of the Spiking Neural Network and analyzed the typical structure of Spiking Neural Network in three aspects,including the learning mechanism,the neuron model and the inhibition mechanism.Then,this paper analyzed the requirements of the design and the problems to be solved of the neural network processor in the paper and put forward solutions,this paper elaborated the whole structure and design scheme as well.Besides,the circuit design of each main function module was explained,and the simulation of each module and the whole system was tested.Finally,in order to verify the application performance of the design,we analyzed a typical SNN algorithm for engineering-improved and applied it to the hardwareplatform.With the accuracy is guaranteed,it improved the efficiency and speed,greatly saved computing resources.In addition,this paper applied the improved algorithm on the SNN processor to verify its performance with the MNIST data set.