Quantized And Adaptive Memristor-based Cellular Neural Networks And System Applications

Cellular Neural Network(CeNN)is a locally interconnected neural network structure that combines the advantages of both cellular automata and artificial neural networks.It has efficient parallel processing capabilities,and its structure is simple and reliable,and it is easy to implement on the hardware level.Therefore,it is widely used in image processing,signal processing and other fields.However,the current CeNN has problems such as insufficient adaptive template design and difficulty in large-scale circuit implementation,and it is difficult to meet the application of complex real-time signal processing and end-side intelligent computing.Therefore,the quantized and adaptive memristor-based CeNN is proposed in this paper.Based on the combination of software and hardware,the fourth type of memristor is introduced to realize the hardware structure.Memristor makes full use of the non-volatile,variable resistance and nanoscale properties of memristor,can realize the adaptive adjustment of the weights of the CeNN template and the large-scale integration and programmability performed on the hardware level.Furthermore,in order to analyze the performance of QA-m CeNN,this paper uses edge extraction and image segmentation in image processing,feature extraction in data analysis as application scenarios to verify and exploration.A large number of comparative experiments and quantitative and qualitative analysis of experimental results are carried out to objectively prove the reality of the project plan.performance.The content of this article is mainly divided into the following parts:At first,in order to solving the defects of CMOS in CeNN,the memristor is introduced.As we all know,memristor is a fourth type device,which has attracted attention because of the unique nano-scale properties,variable resistance,and nonvolatility.In this paper,the memristor is used as the basic circuit element,and the memristor cross-array and the memristive bridge circuit are respectively used as the hardware basis for the realization of CeNN,and a Memristor-based Cellular neural network(m CeNN)is constructed.On the one hand,the artificial synapse structure based on the memristor can realize the programmability of CeNN and provide a basis for the realization of adaptability;On the other hand,the CeNN with nano-level properties can realize the integration of large-scale hardware,further reducing the circuit size.Based on these characteristics,this paper introduces the mathematical model of m CeNN and two circuit structures based on memristor in detail,derives the resistance calculation of the memristive cross circuit,accurately,and analyzes the stability of m CeNN based on the Lyapunov equation.In the next,in order to design the adaptive algorithm of m CeNN template and improve the hardware friendliness of the software algorithm to a certain extent,this paper use a new heuristic optimization algorithm that combines neural network incremental quantization(INQ).Heuristic optimization is an intelligent algorithm that based on actual problems,which can optimize the problem through global search,while,network compression can greatly reduce the severe demands of the neural network on the energy and resource consumption through the combination of clustering,coding,quantization,and sparseness,thereby improving the friendliness of the algorithm at the hardware level.On the one hand,this article uses optimization algorithms to complete the weight training,which can enable the template to be adaptively adjusted according to image characteristics or application scenarios.On the other hand,through the network compression scheme,this paper can fully compress the network weights.While realizing the adaptive adjustment of the template weights,it improves the hardware friendliness of the algorithm with extremely low quantization loss,and realizes the collaborative design of software and hardware.To prove the effectiveness of QA-m CeNN,this article takes the defects of the image processing system and the electronic nose system as the breakthrough point,makes the combination and application of the algorithm proposed in this paper.In the image processing system,the hardware-friendly adaptive image processing(edge extraction and image segmentation)based on QA-m CeNN has been designed.At the same time,in order to improve the stability,robustness and accuracy of the results,the concept of non-linearity has been introduced on the basis of QA-m CeNN,and realizes the non-linear template design of QA-m CeNN,thereby further broadening the application scope of QA-m CeNN(conducive to the realization of complex image processing).Next,in order to verify the effectiveness of the image processing in this article,we performed a large number of edge extraction and image segmentation comparison experiments on different images,and used a variety of evaluation indicators for objective quantitative analysis,and carried out deviations in different situations.The simulation of the analysis experiment confirmed the robustness,stability,effectiveness and accuracy of the image processing scheme in this paper.In the electronic nose system,in order to design a lightweight and high-accuracy gas classification method,we designed a new type of electronic nose data analysis system implementation plan from the selection of gas sensors and data collection as a starting point.Making the combination of QA-m CeNN to realize the hardware-friendly adaptive feature extraction of data sets,adaptive adjustment and training of classifier structure.And then,in order to prove the effectiveness of electronic system,the feature extraction experiment and classifier experiment comparison of different methods are performed on the same data set,and the experimental results are analyzed from multiple angles(including the ROC curve analysis),proves the effective performance of the electronic nose system proposed from this paper in feature extraction and gas data classification,and provides a new idea for the research in the field of edge computing and end-to-side intelligence.