Research On Interpretable Fuzzy System Modeling Method Integrating Deep Feature And Fuzzy Inference

Due to the excellent feature learning ability,deep learning can learn the neural representation of massive data,but it is still facing with a range of challenges with the complex neural network structure,such as the poor interpretability,and nonadjustable prediction result.The upside of fuzzy system is good interpretability which can imitate human inference process,and the prediction result can be corrected by adding additional fuzzy rules,but it also has some limitations,such as optimizing parameters and dealing with high-dimensional data.Therefore,the novel technology that integrates feature learning ability of deep learning and interpretability ability of fuzzy system has become a hot research topic.The relevant researches in this paper mainly focus on the following aspects:1)A novel TSK(Takagi-Sugeno-Kang)fuzzy classifier based on enhanced deep feature(EDTSK-FC)is proposed.ED-TSK-FC adopts the 1D-CNN(One-Dimension Convolutional Neural Network)to extract deep feature and label information.As the antecedent-consequent variable of EDTSK-FC,label information is expanded to deep feature space to generate the enhanced deep feature.FCM(Fuzzy C-means Algorithm)is employed to partition the fuzzy space,the ridge regression extreme learning algorithm is used to solve the consequent parameters of fuzzy rules quickly.EDTSK-FC also formulates a cheap strategy to decrease training time.Experiments show that this method provides an excellent performance on the Bonn epilepsy dataset in terms of classification performance,training time and interpretability.2)To improve the classification performance and interpretability,a novel CNN based deep TSK fuzzy classifier(CNN-TSK-FC)is proposed.In our proposed architecture,it mainly composes of a CNN-based feature learning module and a TSK-type knowledge inference module.From the perspective of fuzzy rule,CNN-TSK-FC is essentially a novel deep TSK fuzzy classifier,it really alleviates the rule explosion problem by compact deep features and improves its interpretability via using original features in the consequent part.Experiments show that this method provides an excellent performance on the electroencephalography(EEG)based boston hospital children's epilepsy dataset(CHB-MIT dataset)in terms of accuracy,sensitivity,specificity and interpretability.3)Due to the insufficient knowledge transferring of ED-TSK-FC and CNN-TSK-FC,a CNN based born-again TSK fuzzy classifier denoted as CNNBa TSK is proposed.With the inherent advantage of fuzzy rule,CNNBa TSK has the capability to express the dark knowledge(soft label information)acquired from the CNN with an interpretable manner.Specifically,soft label information is partitioned into five fixed antecedent fuzzy space.The centers of each soft label information in different fuzzy rules are 0,0.25,0.5,0.75,1,which have corresponding linguistic explanations: very low,low,medium,high,very high.CNNBa TSK also provides a new perspective of knowledge distillation by using a non-iterative learning method(i.e.,Least Learning Machine with Knowledge Distillation,LLM-KD)to train the proposed fuzzy classifier of CNNBa TSK.Experiments on the benchmark datasets and the CHB-MIT dataset demonstrate that CNNBa TSK can improve the classification performance and provide the model interpretability simultaneously.