Convolutional fuzzy modules stacked deep residual system with application to classification problems

Recent years have witnessed tremendous efforts devoted to investigating various data-driven methods, but how to build deep fuzzy models with good interpretability, high-precision, and well generalization ability remains a huge challenge, especially when facing complex, high-dimensional, and strong-nonlinear characteristics in the classification problems. Integrating both the advantages of convolutional neural networks and fuzzy inference method, this paper proposes a deep residual system by stacking the convolutional fuzzy modules (CFM-DRS), which achieves excellent performance with four mechanisms. Firstly, this study designs a new convolutional fuzzy module (CFM), which can comprehensively extract features from datasets with the convolutional operations, and then classify them through the corresponding sub-fuzzy-inference-modules (s-FIM). It is also the foundation of the other three mechanisms. Furthermore, each s-FIM employs the fuzzy C-means algorithm to identify the distribution patterns of features. It not only establishes the inference relationship between the features and output values in an interpretable manner, but also alleviates the problem of rule explosion. In addition, to reduce the impact of outliers and redundancy information on the overall performance, this study adopts the regularization optimization strategy to punish the parameters and prunes the s-FIMs based on their significant contributions. Besides, the utilization of the residual approximation mechanism in the deep framework is beneficial for learning new features and further improving the model’s accuracy. The proposed CFM-DRS is applied to several classification problems. Extensive experiments on different benchmark and real-world datasets demonstrate that the proposed CFM-DRS has a better classification performance compared to several state-of-the-art methods.