基于分类特征约束变分伪样本生成器的类增量学习

针对神经网络模型进行类增量训练时产生的灾难性遗忘问题,提出一种基于分类特征约束变分伪样本生成器的类增量学习方法.首先,通过构造伪样本生成器记忆旧类样本来训练新的分类器及新的伪样本生成器.伪样本生成器以变分自编码器为基础,用分类特征进行约束,使生成的样本更好地保留旧类在分类器上的性能.然后,用旧分类器的输出作为伪样本的精馏标签,进一步保留从旧类获得的知识.最后,为了平衡旧类样本的生成数量,采用基于分类器分数的伪样本选择,在保持每个旧类伪样本数量平衡的前提下选择一些更具代表性的旧类伪样本.在MNIST、FASHION、E-MNIST和SVHN数据集上的实验结果表明,所提出的方法能有效减少灾难性遗忘的影响,提高图像的分类精度.

Class incremental learning based on variational pseudo-sample generator with classification feature constraints

Aiming at the catastrophic forgetting problem caused by the class incremental training of neural network models, a class incremental learning method based on a variational pseudo-sample generator with classification feature constraints is proposed. Firstly, a new classifier and a new pseudo sample generator are trained by constructing a pseudo-sample generator to memorize old class samples. The pseudo sample generator is based on the variational autoencoder and uses classification features to constrain the generated samples to better retain the performance of the old class on the classifier. Then, the output of the old classifier is used as the distillation label of the pseudo sample to further retain the knowledge obtained from the old class. Finally, in order to balance the number of samples generated by the old class, pseudo sample selection based on the score of the classifier can be used to select some more representative samples of the old class while maintaining the balance of the number of pseudo samples of each old class. Experimental results on MNIST, FASHION, E-MNIST and SVHN datasets show that the proposed method can effectively reduce the impact of catastrophic forgetting and improve the accuracy of image classification.