Continuous Attractors Of Coupled Neural Networks

With the advancement of technology,humans have a clearer understanding of the brain,and have established many models of cognitive brain.As one of the models,neural network has achieved good results in many aspects.Coupled neural network is a very important part of neural network model,which has two forms.One is a parallel network,which combines different neural networks in parallel,such as modular neural network,distributed neural network;the other is a series neural network,such as auto-encoder model network,which can be regarded as a series between levels in the mathematical model.One of the most important aspects of neural network is the study of attractors.In this thesis,the continuous attractors of coupled neural network are studied and discussed from the theoretical aspect.The research contents are mainly divided into the following parts:Firstly,different kinds of coupled neural networks based on coupling coefficients and activation function are studied in this thesis.We divide the common combinations of coupled neural networks,and get the condition for continuous attractors in coupled recurrent neural networks and verify the theory through simulation.Secondly,the auto-encoder neural network of threshold linear function is studied in this thesis.In this part,we study the existence condition of attractor of coupled neural network,combine auto-encoder network with continuous attractor,study the existence form of continuous attractor.Then we extend the existing results,and theoretically deduce the input and external input model of auto-encoder network,and verify the theoretical deduction through simulation of the correctness of the derivation.Finally,we combine auto-encoder model with practical application.We combine the business logic of the scorecard,use automatic coding to reconstruct the data of related variables,and combine the automatic coding model with logistic regression,and achieve good results.