Incremental learning with open set based discrimination enhancement

摘要

Multi-Class Incremental Learning (MCIL) is an important task in deep learning that aims to solve real world problems of allowing models to continuously learn new concepts without forgetting old ones. However, MCIL typically suffers from catastrophic forgetting. In addition, it is difficult to implement many MCIL algorithms, which require to store data associated with the number of existing classes, on edge devices with limited memory. We believe that the data imbalance between the old and new classes and the utilization of old classes in the incremental process, which increases the memory requirements, are the main reasons for the above problems. Hence, we propose a novel approach, called ‘Open Set Incremental Learning(OSIL)’, to preserve the information of old classes, without storing any of their data, while making the classifier progressively learn the new classes and forget less information about the old classes. In OSIL, we present an open-set incremental strategy, which uses enhanced discrimination combined with open-set identification to overcome the imbalance between old classes and new classes. Meanwhile we demonstrate that attention mechanisms help to address catastrophic forgetting. The proposed method is evaluated on CIFAR-100 and ImageNet-100 under various settings. Experimental results show that OSIL effectively alleviates catastrophic forgetting and significantly outperforms state-of-the-art methods.