A Configurable Modularization Hardware Implementation Method For SOM Neural Network Based On FPGA

Learning is an important function of the body's brain.The SOM neural network model proposed by Prof.Kohonen in 1981 is widely used in the simulation and research of brain self-organization and self-learning functions.At the same time,it has also achieved great success in data reduction,clustering,pattern recognition,data mining and image processing.At present,the implementation and application of artificial neural networks mostly use computer-based software simulation methods.This method cannot really take the advantages of parallel processing,distributed storage,strong self-adaptability and strong robustness of neural network.Therefore,the hardware implementation of neural networks has gradually become a research hotspot.At present,the hardware implementation of the artificial neural network is mainly based on Very Large Scale Integration(VLSI)technology.Among them,Field Programmable Gate Array(FPGA)is widely used in hardware implementation of neural network because of its rich hardware programmable logic resources,flexible configuration,short development cycle,and parallel computing.Most existing FPGA-based SOM neural network hardware implementation methods are oriented to specific applications,the network size is fixed,the data format is single,and the internal implementation structure cannot be configured on demand.It does not have versatility and it is difficult to achieve high-performance artificial neural networks.Therefore,there is an urgent need for a universal SOM neural network hardware implementation method based on FPGA,by which the network size,data range and precision,and an arithmetic structure that can be configured on demand.This paper presents a hardware implementation method of configurable and modular SOM neural network based on FPGA.Firstly,the model structure,learning principle and algorithm of the SOM neural network is integrated with the conditions and limitations of the FPGA hardware implementation.The network model is divided into modules with relatively independent structure and function;secondly,the hardware description language VHDL is used to describe each module and complete the digitalization of the module to form a common module library;Finally,in practical applications,according to the requirements,the target hardware of the SOM neural network can be constructed through the combination of each module.When constructing a network,flexible configuration of network size,data range,precision,and operation structure can be accomplished by setting and transferring common parameters.To verify the effectiveness of this method,a hardware SOM neural network test system for digital identification was constructed.The system consists of a host computer setup interface based on MATLAB GUI and a lower machine based on FPGA development board.Tests of network functions and performance show that the hardware implementation method of SOM neural network proposed in this paper has flexible network scale,data format,and configurable ability of operation structure,and can achieve optimal design of resource occupation and operation speed.The test results of the system running speed show that,compared to the software implementation method,the SOM neural network implemented by the hardware has a high running speed and can meet the application requirements of the high-speed and miniaturized intelligent information processing system.This paper provides a new method for the application of SOM neural network,and it can also provide a valuable reference for the hardware implementation of other types of artificial neural networks.