Research On Tactile Data Classification Algorithm Based On Spiking Neural Networ

With the development of artificial intelligence technology,intelligent robots are widely used in service industry,medical and other fields.However,most robots can still only perceive the external environment through visual and auditory means,and are unable to sense and process tactile information about objects.Therefore,expanding the tactile perception capability of intelligent robots can enhance the autonomous perception capability of robots,which can promote the application of robots in hazardous environments,thus improving the intelligence level and task execution capability of robots.Tactile data is the basis of tactile perception,which is used to perceive features such as hardness,texture,and shape of objects,therefore,the study of classification algorithm based on tactile data is a necessary step to achieve the tactile perception.Aiming at the spatio-temporal characteristics of tactile data,we selected spiking neural networks,which has rich spatio-temporal dynamics and the advantage of processing event-based scenes,as the basic method to build a classification algorithm for tactile data based on spiking neural networks.The main study of this paper is as follows.(1)The unique spiking activation mechanism in spiking neural networks makes it impossible to directly use the gradient descent method in back propagation to update network parameters,so that it cannot accomplish the classification task of tactile data.To address this problem,this paper proposes a spiking neural network based on gradient approximation method for tactile data classification algorithm.By selecting four different back-propagation approximation functions,the gradient approximation method is used to overcome the discrete non-differentiable problem of activation function,and a network model based on spiking neural networks is constructed to complete the classification task of tactile data.In this way,the foundation for the subsequent algorithm and model optimization is laid.(2)The acquisition of tactile data requires specialized equipment,however,the existing tactile spike dataset has few data samples,which can lead to overfitting problems in the constructed model.Aiming at this problem,this paper proposes an overfitting optimization method for classification models of tactile data with regularization constraints.Considering both the loss function and the network structure of the model,the Dropout regularization method,the L2 regularization method and the cosine annealing method are introduced to alleviate the model overfitting problem and improve the model performance.(3)The regularization method helps the model to improve the prediction accuracy,but brings the increase of the model convergence value,which leads to the model underconvergence problem.Addressing this problem,an improved under-convergence optimization method for the classification model of tactile data under membrane potential representation is proposed in this paper.The spatio-temporal dynamics equations of neurons in the impulsive neural network are solved by Euler’s method,in order to improve the expression of the membrane potential of impulsive neurons,and the relationship between the loss function and the weight and bias of the model in back propagation is updated.In addition,the time overhead is introduced as one of the evaluation metrics of the algorithm.The method is experimentally proven to be effective in reducing the convergence value of the model and speeding up the convergence rate,as well as reducing the time spent on training and testing of the model.