Broad Learning System for Class Incremental Learning

Abstract

The large-scale image datasets such as ImageNet and open-ended photo websites are revealing new challenges to image classification that were not apparent in smaller and fixed sets. In particular, how to handle the dynamically growing datasets efficiently, where not only the amount of training data but also the number of classes increases over time, remains an unexplored problem. In this challenging setting, we study how to employ the Broad Learning System (BLS) to deal with the incremental increases of sample classes in datasets. We first determine whether a new batch of samples is from certain seen classes or unseen classes by two methods based on the statistics theory. Then, the incremental training algorithms are developed to classify unlearned data from new classes. We test our class incremental learning algorithms of BLS on some representative datasets such as Cifar-100. The experimental results show that our methods can accurately distinguish seen and unseen classes. And the comparison study demonstrates that our method can be extended to 100 classes on Cifar-100 with an acceptable loss of accuracy and performs better compared to other similar class incremental algorithms such as iCaRL.