Prototypes Sampling Mechanism for Class Incremental Learning

Incremental learning aims to alleviate the catastrophic forgetting problem of deep neural networks during sequential data stream. This is even more challenging when old unavailable, since system can only be trained under supervision current data. To address this problem, we proposed a prototype sampling mechanism based on K-means clustering method. On one hand, use pick out class-representative prototypes for each class. During incremental stages, and features from are together maintain distinction balance between new classes. other attach mask loss function cosine similarity Which further enhances discrimination classes compared naive knowledge distillation schemes. Extensive experiments conducted three benchmark datasets including CIFAR100, Tiny-ImageNet vggface2 verified effectiveness advantages our Specifically, improved class performance by 1.6%, 1.2% 1.7% respectively.