Incremental classifier learning based on PEDCC-loss and cosine distance

Traditionally, the performance of deep convolutional neural networks relies on a large number labeled datasets in advance. However, real-world applications, training data are not collected at once, so an algorithm which can deal with continuous incoming is needed. This learning method called incremental learning, whose main problem catastrophic forgetting. Neural network will perform badly old classes after for new classes. To solve this problem, paper proposes integrated approach based PEDCC-loss, and cosine distance between sample’s output feature PEDCC (predefined evenly distributed class centroids). Old knowledge learned by stored separately. During each network, PEDCC-Loss used to constrain their corresponding pre-defined center. Meanwhile, retained different rates across retention mode samples discussed. In test phase, final prediction determined distances features all networks. Our experiments EMNIST, CIFAR100 TinyImageNet show that our learn incrementally without quickly failing performance. Compared some existing algorithms, such as Hou, iCaRL finetune, has better