Prototypes Sampling Mechanism for Class Incremental Learning

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