Recurrent Neural Network Based On Memristive Activation Function And Its Associative Memory

The memristor is a nonlinear dynamic nanometer device.With a good many characteristics such as low power consumption,fast speed and easy to obtain a integration with high density,it is of a great significance for hardware circuit.However,for the research of memristor in the neural network field,the researchers use the synaptic characteristics of the memristor to realize the synapses of the neural network,other than other properities.Therefore,this article,based on the method that the nonlinear characteristic of the flux-controlled memristor can be used to achieve memristive activation function circuit,with low power consumption and a good extendibility,proposes a recurrent neural network based on memristive activation function(M-RNN)and applies it to associative memory.Then,by extending the single-layer M-RNN to multi-layer M-RNN,it further proposes a multi-layer recurrent neural network based on memristive activation function(MM-RNN),and testifies its stability and validity with mathematical analysis.Moreover,by using MM-RNN,the associative memory of dynamic image can be achieved.The main contents are as follows:First of all,the flux-controlled memristor is connected in parallel,and the nonlinear characteristic between magnetic flux and charge is used to realize memristive activation function.Based on the theories above,thus,it can provide a new and effective method to realize the large-scale integration of recurrent neural network circuits from the angle of other characteristics,instead of memristor synapse.Secondly,different from the previous researches of recurrent neural network,this article applies the memristive activation function to the traditional recurrent neural network,adds the matrix transfer function ?(t),and proposes a M-RNN.According to the relationship between ?(t)and input image P,associative memory of M-RNN can be classified into static image associative memory and dynamic image associative memory for simulation analysis.The two-value and three-value static images of autoassociative memory and hetero-associative memory can be achieved by MATLAB simulation.In addition,by adding a matrix transfer function ?(t),M-RNN can also realize simple dynamic image associative memory.Last but not least,extending the single-layer M-RNN to multi-layer M-RNN,a new MM-RNN is proposed.At the same time,it also increases the complexity of the analysis and even affects the convergence of the whole network.Therefore,the stability of MMRNN is proved by mathematical theory.Then,through the associative memory experimental simulation,the stability and validity of MM-RNN turns out to be positive.With the background and the foreground of the target image acting as input of the twolayer network respectively,it can not only realize the static image of associative memory,but achieve the dynamic image of associative memory,which proves that MM-RNN has a more powerful ability of processing information than M-RNN.In this way,this research can provide an innovative choice for exploring associative memory of video images,to some degree.