Design And Implementation Of Neural Network Basic Modeling Framework Based On MSVL

With the continuous upgrading of computer software and hardware,artificial intelligence,especially neural network systems,has been rapidly developed and applied in various fields of daily lives.However,every misjudgment of the neural network may bring about security problems.Neural network is an optimization model based on mathematical statistics.Its purpose is to use approximate functions to efficiently process input-output relationships that are difficult to describe in the physical world.This means that neural networks have natural uncertainties for unknown inputs;while formal method is a modeling and verification method based on rigorous mathematical logic that is suitable for describing software and hardware systems.Its purpose is to improve the reliability of software and hardware design.It is a typical deterministic result-oriented method.Applying formal methods to the modeling and verification process of neural networks will help to improve the uncertain relationship of neural networks and at the same time increase the credibility of neural network output results.This thesis is based on Modeling,Simulation and Verification Language(MSVL),to achieve a formal neural network system model framework that provides the necessary guarantee for the formal verification of the neural network system.This thesis proposes a basic neural network modeling framework based on MSVL,which supports modeling of a general type of neural network,back propagation neural network.First of all,the model framework forms a design idea of hierarchical neural network structure and flow-based neural network behavior in accordance with the basic design principles.Specifically,the hierarchical neural network is designed as a basic hierarchical data structure and the process-based neural network behavior is designed as a basic data operation.Then we use the design as a blueprint to formalize the hierarchical structure into the MSVL structure,while we formalize the basic data operations into the MSVL function.On this basis,the construction of the neural network process operation is formalized step by step into several sets of neural network construction functions,forming a prototype of a formal function library for neural network system modeling that can be compiled and run under the MSVL compiler.Based on this library,a back-propagation neural network model can be established,and the model can be constructed and trained according to the user's training requirements.Finally,for the public handwritten font recognition image data set,through several sets of comparative experiments,the correctness of the model framework is analyzed,and the algorithm efficiency and running time are optimized.The results show that these models have good performance in terms of training and prediction indicators.