Fuzzy Kernel Net

The convolutional neural network is a commonly used neural network architecture in the field of deep learning.However,as neural network structures gradually move towards large parameter models,the required computational complexity for model training continues to increase.This leads to a significant increase in the computational cost of model training,which severely limits the application and development of deep learning models.Effectively reducing the computational complexity of neural network models while ensuring the training effectiveness can help solve this problem.Therefore,this paper proposes a neural network convolution structure network that utilizes fuzzy rules to improve the convolution process in CNN,called the Fuzzy Kernel Structure Network(FKSN).The main contributions of this paper are as follows:Firstly,based on existing kernel structures,five different fuzzy kernels were designed using fuzzy rules.By combining the convolution kernel and fuzzy rules in CNN,the fuzzy kernel can achieve arbitrary size with extremely low computational cost.Then,based on the fuzzy kernel,a fuzzy convolution layer was constructed by combining the traditional convolution process,which can produce convolution effects similar to the conventional convolution used in CNN.Finally,the fuzzy convolution layer was combined with popular CNN architectures to construct various forms of FKSN,and their performance was experimentally tested.After conducting comparative tests on different FKSNs on benchmark datasets,the test results showed that the proposed FKSNs exhibited better model performance in terms of parameter count during training,suppression of overfitting,and model accuracy compared to CNNs that use ordinary convolution.