New Fuzzy Deep Randomized Networks Boost Classification Robustness

Current state-of-the-art deep randomized neural networks, such as dRVFL and edRVFL, often treat all training samples uniformly. This approach limits their robustness and effectiveness, especially when dealing with real-world datasets that contain noise and outliers. Furthermore, the propagation of contaminated features through hidden layers can negatively impact the models' decision-making capabilities. To address these limitations, a new framework introduces intuitionistic fuzzy dRVFL (IF-dRVFL) and intuitionistic fuzzy edRVFL (IF-edRVFL). These proposed models integrate intuitionistic fuzzy theory to leverage sample neighborhood information within the kernel space. They achieve this by jointly considering both membership and non-membership degrees for each sample. Membership degrees are calculated based on a sample's distance from its respective class centroid, while non-membership degrees quantify the sample's heterogeneity within its local neighborhood. These measures are then used to assign adaptive weights to training samples, allowing the models to effectively distinguish between clean, noisy, and outlier data points. Extensive experiments on UCI and KEEL benchmark datasets, including those with Gaussian noise, demonstrate that IF-dRVFL and IF-edRVFL consistently outperform existing fuzzy and non-fuzzy approaches, significantly enhancing model robustness.