Abstract:The broad learning system (BLS) has recently been applied in numerous fields. However, it is mainly a supervised learning system and thus not suitable for specific practical applications with a mixture of labeled and unlabeled data. Despite a manifold regularization-based semi-supervised BLS, its performance still requires improvement, because its assumption is not always applicable. Therefore, this article proposes an incremental-self-training-guided semi-supervised BLS (ISTSS-BLS). Distinctive to traditional self-training, where all unlabeled data are labeled simultaneously, incremental self-training (IST) obtains unlabeled data incrementally from an established sorted list based on the distance between the data and their cluster center. During iterative learning, a small portion of labeled data is first used to train BLS. The system recursively self-updates its structure and meta-parameters using: 1) the double-restricted mechanism and 2) the dynamic neuron-incremental mechanism. The double-restricted mechanism is beneficial to preventing the introduction of incorrect pseudo-labeled samples, and the dynamic neuron-incremental mechanism guides the self-updating of the network structure effectively based on the training accuracy of the labeled data. These strategies guarantee a parsimonious model during the update. Besides, a novel metric, the accuracy-time ratio (A/T), is proposed to evaluate the model's performance comprehensively regarding time and accuracy. In experimental verifications, ISTSS-BLS performs outstandingly on 11 datasets. Specifically, the IST is compared with the traditional one on three scales data, saving up to 52.02% learning time. In addition, ISTSS-BLS is compared with different state-of-the-art alternatives, and all results indicate that it possesses significant advantages in performance.

Broad Learning System for Class Incremental Learning

ConvBLS: An Effective and Efficient Incremental Convolutional Broad Learning System for Image Classification

A Broad Neural Network Structure for Class Incremental Learning.

Broad Learning System: An Effective and Efficient Incremental Learning System Without the Need for Deep Architecture

An Incremental Broad Learning Approach for Semi-Supervised Classification

Incremental Bayesian broad learning system and its industrial application

A Hybrid Recursive Implementation of Broad Learning with Incremental Features

Analysis and Variants of Broad Learning System

Class-Incremental Learning: Survey and Performance Evaluation on Image Classification

Class-Incremental Learning: A Survey

Stacking Multi-View Broad Learning System with Residual Structures for Classification

Robust Incremental Broad Learning System for Data Streams of Uncertain Scale

Large Scale Incremental Learning

Stacked Broad Learning System: From Incremental Flatted Structure to Deep Model

An Incremental-Self-Training-Guided Semi-Supervised Broad Learning System

Discriminative elastic-net broad learning systems for visual classification

Broad learning system: Feature extraction based on K-means clustering algorithm

Learning to Classify With Incremental New Class

Broad Learning With Reinforcement Learning Signal Feedback: Theory and Applications

Error-Driven Incremental Learning in Deep Convolutional Neural Network for Large-Scale Image Classification

Class-Incremental Learning with Strong Pre-trained Models