Membership Functional Analysis for Nonparametric Deep Models of Image Features

The application of fuzzy theory to deep learning is limited 1) under the realm of deep neural networks; 2) to the parametric form of modeling; and 3) relying on gradient-descent based numerical algorithms for optimization because of lack of analytical solutions. This study fills this gap by providing an analytical nonparametric deep modeling solution based on the mathematical analysis of membership functions assigned to model variables. The nonparametric approach is based on the concept of representing the unknown mappings (between input and output variables) through a fuzzy set with Student-t type membership function such that the dimension of membership function increases with an increasing data size. This concept of function representation is referred to as “Student-t fuzzy-mapping” in this study. The most significant feature of this study is to analytically derive the mathematical expressions for
membership functions (which quantify uncertainties regarding the values of variables) using variational optimization such that the degree-of-belongingness of given data to the considered data-model is maximized. The study focuses on the modeling of image features where a layer of the deep-model first projects the feature vector onto a lower dimensional subspace and then construct the output feature vector through Student-t fuzzy-mappings. Numerous
image classification experiments are provided to support the proposed approach.

Related Publications

• M. Kumar and B. Freudenthaler, “Fuzzy Membership Functional Analysis for Nonparametric Deep Models of Image Features,” IEEE Transactions on Fuzzy Systems, https://doi.org/10.1109/TFUZZ.2019.2950636.