Hybrid Deep-ANFN Model for Dimensionality Reduction and Classification

Abstract

This paper proposes a novel method to enhance classification performance and the reduction of high-dimensional feature maps by fusing deep learning with adaptive neuro-fuzzy network (ANFN). Although deep learning can be used to generate input data as high-dimensional nonlinear features, thereby achieving high classification performance, this method challenges remain in interpreting data with uncertain meanings. Because fuzziness is a characteristic that can be used to process and interpret such data, research is being actively conducted to combine the advantages of fuzziness and deep learning. In conventional fusion methods, high-dimensional feature maps are reduced via principal component analysis (PCA), a linear dimension reduction technique. We propose a fuzzy feature selection (FFS) scheme that maintains nonlinear patterns, and the One-vs-Rest Deep-ANFN Fusion model (ORDAF) to identify feature channels useful for each class. Experiments show that the proposed method achieves comparable or superior performance to that of the conventional method with only 5–20 feature channels per class among the high-dimensional features of deep learning. To evaluate the performance of ORDAF through feature selection optimization, the Defungi, Kaggle images, and CIFAR-10 datasets were used for channels specific to each class, with fewer feature channels than conventional deep-learning-based classification. In this study, classification was facilitated by removing unnecessary feature channels from the model, and a more direct interpretation was obtained by checking specialized channels. In subsequent studies, these methods can be used to develop a highly explainable technology that determines which area of the original data corresponds to each selected channel.

Keywords

• Artificial intelligence
• explainable AI
• deep learning
• fuzzy inference systems
• adaptive neural fuzzy inference systems
• feature selection