Research On Spiking Neural Network Model Based On Visual Mechanism

Spiking Neural Network(SNN)takes spiking neurons as the basic unit to simulate the information transmission form between biological neurons.It is a new generation of artificial neural networks inspired by biology.The spikes emitted by biological neurons are sparse in the temporal and spatial domains,and the spikes are event-driven.Combined with the local learning rules of the brain,SNN is biologically more interpretable than traditional artificial neural networks.At the same time,neuroscience research shows that the SNN model using spike coding can obtain more information and has stronger computing power.At present,more and more attention has been paid to the related research of SNN,such as spike encoding scheme,network structure optimization,etc.Different models and algorithms have also been proposed.However,due to its inherent mechanism and other problems,the performance of SNN in pattern recognition tasks is far from the level that can compete with biological vision systems,and confusion in different patterns is often accompanied in image recognition.In addition,how to design an efficient learning algorithm for SNN,which topology is more effective,and how to apply it to more complex visual tasks(such as object detection)are still important issues in this research field.The information processing mechanism of biological vision system is the key to break through the bottleneck of computer vision development.Traditional artificial neural networks focus on the research of attention mechanism.If we want to reach the level equivalent to the visual system,SNN needs to combine more appropriate visual mechanisms to improve the network performance.In view of the above reasons,this paper,from the perspective of improving the performance of the model,starts from simulating the working mechanism of the brain visual system,and carries out research on the SNN model and its application based on the visual mechanism.The research contents are summarized as follows:1.Aiming at the insufficient of the SNN’s ability to extract features of input patterns,an SNN based on visual saliency and two spike encoding schemes are proposed with the help of the information processing mechanism of the visual system.At the same time,the structure of SNN is further optimized using the FOA to improve the recognition performance.By verifying the performance of the proposed method on different types of image datasets,the results show that the network is more effective for tasks with fewer image categories.The time encoding scheme based on information entropy has better performance than the linear time encoding scheme,and the recognition accuracy can be further improved by combining saliency calculation and FOA optimized SNN.2.Aiming at the poor performance of SNN in multi-classification tasks and confusion in different patterns,an ensemble unsupervised deep SNN model based on visual lateral inhibition is proposed.According to the brain mechanism of primates,excitatory neurons and inhibitory neurons compose an SNN network with a ratio of 8:2.The input pattern is encoded by rank order,and the visual features are extracted by unsupervised learning using STDP learning rules in the convolution layer.The dynamic spike threshold form is adopted by spiking neuron model,and the lateral inhibition mechanism is combined in the classification stage to suppress the nonfiring neurons to produce distinguishable results,so as to solve the problem of pattern aliasing.To improve the performance of SNN,an ensemble SNN architecture based on voting method is adopted,and transfer learning is used to avoid repeated training of SNN when solving different tasks.Through testing on several datasets,the results show that the proposed model can obtain correct recognition results for different recognition tasks.3.A parallel convolution SNN model combined with adaptive lateral inhibition is proposed to solve the problem of the shortage of lateral inhibition computation and the diversity of neurons.Butterworth filter is added to the calculation of adaptive lateral inhibition to effectively suppress the noise information,and the lateral inhibition coefficient is adaptively adjusted exponentially according to the information entropy.The network is composed of evolutionary LIF neurons.With the training of SNN,the membrane potential parameters can be dynamically adjusted according to the network output error to maintain the diversity of neurons in the network.The parallel convolutional SNN is adopted to make the network extract input features better.The proposed method is verified on static datasets and neuromorphic datasets,and it is applied to breast tumor recognition.4.Aiming at the more difficult visual task,i.e.object detection,an object detection network based on SNN is proposed.The network takes YOLO framework as the carrier,and adopts the conversion method from DNN to SNN to train the network parameters,which can avoid the failure of gradient descent algorithm to train SNN directly and maintain the detection performance of the network.The model combines feature pyramid and visual saliency to improve the accuracy of network target detection.Through the experimental verification on the breast tumor detection tasks,the experiments show that the object detection network based on SNN can obtain a satisfactory detection results.