Learning with randomized neural networks for classification problems

Abstract

Randomized neural networks (RdNNs) have shown their strength in classification and regression problems. RdNNs with less computational cost and good generalization performance are highly desirable machine learning models. In RdNNs such as random vector functional link (RVFL) neural network, some parameters are kept fixed (during training), either in a stochastic or a deterministic way, and rest parameters are optimized via closed form or iterative methods. RVFL with direct links is a special randomized network. By using a closed form solution approach, RVFL avoids concerns that back-propagation trained networks experience, such as the local minima problem, sluggish convergence, and sensitivity to learning rate setting. This thesis aims to contribute to the evolution of RVFL by developing novel variants of RVFL for classification problems. We give an extensive review of the progress of RVFL, which is useful for beginners as well as professionals. RVFL assumes that all the samples are equally important, however, this may not be true in real world scenarios. To handle this issue, we employ fuzzy and intuitionistic fuzzy theory to reduce the negative influence of the noise/outliers over RVFL’s performance and develop intuitionistic fuzzy RVFL (IFRVFL) and class probability-based fuzzy RVFL (CP-FRVFL) models. However, IFRVFL ignores the geometrical information of the data while calculating the final parameters.