CIBLS-PLS: A Class-Incremental Broad Learning System With Pseudolabel-Guided Stacked Structure

Class-incremental learning (CIL) with the Broad Learning System (BLS) has emerged as a computationally efficient alternative to deep incremental models. However, existing BLS-based CIL methods struggle with complex data distributions and are highly sensitive to hyperparameter tuning, leading to suboptimal knowledge retention. To address these challenges, we propose CIBLS-PLS (Class-Incremental Broad Learning System with Pseudo-Label-Guided Stacked Structure), which enhances knowledge retention and adaptability. Different from traditional stacked BLS, where blocks are strictly chained through input-output dependencies, CIBLS-PLS adopts a more flexible stacking structure, allowing each block to independently contribute to knowledge preservation. Each BLS layer integrates dual-storage modules that retain key features from previous data, while pseudo-labels are generated to facilitate seamless knowledge integration within the stacked residual learning framework. Model parameters are updated efficiently via closed-form ridge regression, significantly reducing computational overhead while maintaining high accuracy. Additionally, to further enhance model generalization, an adaptive scaling mechanism dynamically regulates the contribution of residual blocks, effectively preventing overfitting as the number of blocks increases. This property is rigorously validated through both theoretical analysis and extensive experiments. Results on seven large-scale image datasets demonstrate that CIBLS-PLS achieves state-of-the-art performance in accuracy and knowledge retention while maintaining competitive computational efficiency, paving the way for robust and scalable broad learning-based incremental models.