Central-Diffused Instance Generation Method in Class Incremental Learning

Abstract

Class incremental learning is widely applied in the classification scenarios as the number of classes is usually dynamically changing. Meanwhile, class imbalance learning often occurs simultaneously in class incremental learning when the new class emerges. Previous studies mainly proposed different methods to handle this problem. But these methods focus on classification tasks with a fixed class set and cannot adjust the peripheral contour features of the original instance distribution. As a result, the classification performance degrades seriously in an open dynamic environment, and the synthetic instances are always clustered within the original distribution. In order to solve class imbalance learning effectively in class incremental learning, we propose a Central-diffused Instance Generation Method to generate the instances of minority class as the new class emerging, called CdIGM. The key is to randomly shoot direction vectors of fixed length from the center of new class instances to expand the instance distribution space. The vectors diffuse to form a distribution which is optimized to satisfy properties that produce a multi-classification discriminative classifier with good performance. We conduct the experiments on both artificial data streams with different imbalance rates and real-world ones to compare CdIGM with some other proposed methods, e.g. SMOTE, OPCIL, OB and SDCIL. The experiment results show that CdIGM averagely achieves more than 4.01%, 4.49%, 8.81% and 9.76% performance improvement over SMOTE, OPCIL, OB and SDCIL, respectively, and outperforms in terms of overall and real-time accuracy. Our method is proved to possess the strength of class incremental learning and class imbalance learning with good accuracy and robustness.