Research And Implementation Of Logical Operation Based On Spiking Neural Networks

In the 21 st century,artificial intelligence has become a hot field for all countries to seize development opportunities.With the deepening of artificial intelligence research,special artificial intelligence has achieved many superhuman achievements in some tasks,but it also faces many problems and deficiencies.On the way from special artificial intelligence to general artificial intelligence,brain-inspired intelligence gradually comes into people’s sight and expands the possibility of realizing general artificial intelligence with stronger biological interpretability.As one of the frontier technologies and industrial transformation fields,braininspired intelligence was also included in the Outline of the 14 th Five-Year Plan(2021-2025)for National Economic and Social Development and Vision 2035 of the People’s Republic of China in 2021.The research of brain-inspired intelligence is inseparable from the spiking neural networks,which is called the third-generation neural network.In recent years,the field of spiking neural networks has made exciting progress in many aspects.Spiking neural networks is the basis of building brain-inspired intelligence,logical operation is not only the basis of complex operation,but also the basis of advanced intelligence.The essence of biological neural activity is the process of calculating external input signals and self-excited signals.The logical operation through spiking neural networks has biological rationality and practical application value,which is of great significance for the development of artificial intelligence to general artificial intelligence and the further development of brain-inspired intelligence.This thesis carries out research based on the above,and the specific research content includes:(1)Logical operation paradigm design based on spiking neural networks.In this thesis,a unified logical operation paradigm Logic SNN based on spiking neural networks is designed,including the definition of logic variables and the method of spike coding,the comparison and selection of spike neurons,the design of paradigm network structure and the determination of network training methods.In the context of spiking neural networks,the two logical constants of binary logic variable,logic 0 and logic 1,are defined as two different neurons firing spikes,that is,the two states of each logical variable are represented by two independent neurons.The input encoding method adopts time-domain mutually exclusive code,and the output conforms to the above definition through training.After comparison,LIF neurons were selected as the spiking neurons to construct the paradigm.The paradigm network is designed with four layers in the training stage and three layers in the application stage.The input layer,pattern layer and output layer are hierarchical structures from the front to the back of the application stage.The training stage structure adds a teacher layer based on the above hierarchical structure.The teacher layer is connected with the output layer for network training.STDP learning rule meeting Hebb rule is adopted in the network training method.(2)The realization of logical operation based on the logic operation paradigm of spiking neural networks.To verify the universality,applicability and effect of the designed paradigm,the designed paradigm was verified.Basic logical operations(AND,OR,NOT),some commonly used combinational logical operations(NAND,NOR,XOR,XNOR)and some of the combinational logical networks(rounding logic network of 8421-BCD code,half adder,full adder)are trained and implemented based on the paradigm Logic SNN.All experiments achieve the expected logic functions.And the feasibility,universality,applicability and effect of the paradigm are verified.Among them,the basic logical operation NOT can be implemented by a two-layer network structure containing only the input layer and output layer because of its inverting characteristic.Basic logic operations AND,OR and combinational logical operations NAND,NOR,XOR,and XNOR are trained with the paradigm Logic SNN,and the corresponding logical modules are obtained after training.The six kinds of logical operation modules reflect the unified network structure characteristics of the Logic SNN paradigm,different logical operation modules only have different synaptic weights.The realization of the rounding logic network of 8421-BCD code,half adder,full adder of the combinational logic network makes use of the cascading characteristics of the Logic SNN paradigm.After analyzing the logic functions of the network,the trained logical operation modules are used to construct the corresponding network functions in the way of "building blocks".(3)Research on generalized logic function based on spiking neural networks.According to the definition of the logical function,this thesis extends the general logical function to the generalized logical function.The decision tree,one of the traditional machine learning algorithms,and partial unconditioned reflex behavior of the model organism Caenorhabditis elegans are analyzed.And make it logical.Based on spiking neural networks’ logical operation modules,the decision tree is built to classify the IRIS dataset,and the foraging behavior and avoidance behavior of Caenorhabditis elegans under a single stimulus and multi stimulus environment are simulated.All of these experiments achieve the expected logical functions.This makes logical operations based on spiking neural networks applied to wider areas.