Stacking multi-view broad learning system with residual structures for classification

Abstract

Broad learning system (BLS) is an effective and efficient discriminative learning algorithm, particularly adept at rapid implementation for incremental learning without necessitating substantial computational resources. However, BLS exhibits excellent performance exclusively in low-complexity scenarios, with its classification performance on RGB images being somewhat underwhelming. In this article, a novel BLS model named the stacking multi-view broad learning system with residual structures (RSM-BLS) is proposed. This model integrates the strengths of residual structures, multi-view learning, and transfer learning. Furthermore, the specific architecture of this model and the process of implementing incremental learning are provided. Finally, we evaluate the classification performance of the proposed RSM-BLS on the NOBR dataset, Fashion-MNIST dataset, cifar10 dataset, SVHN dataset, and cifar100 dataset, and conduct ablation studies and incremental structure tests. The experimental results indicate that, in comparison to relevant BLS algorithms and state-of-the-art methods, RSM-BLS exhibits superior performance and generalization capabilities.