Rethinking class orders and transferability in class incremental learning

Chen He, Ruiping Wang, Xilin Chen

Abstract: Class Incremental Learning (CIL), an indispensable ability for open-world applications such as service robots, has received increasing attention in recent years. Although many CIL methods sprouted out, researchers usually adopt default class orders, leaving the characteristics of different class orders less visited. In this paper, we rethink class orders in CIL from the following aspects: first, we show from preliminary studies that class orders do have an impact on the performance, and mainstream episodic memory-based CIL methods generally favor an interleaved way of arranging class orders; then, we interpret the phenomena above with transferability and propose transferability measures of class orders, which are in line with the method performance under different class orders; based on that, we propose a Class Order Search Algorithm (COSA) to obtain an optimal class order by finding which one has almost the highest transferability. Experiments on Group ImageNet and iNaturalist verify the importance of class orders in CIL methods, and demonstrate the effectiveness of our proposed transferability measures and COSA. These findings may help raise more attention to the hardly visited class orders in CIL.