基于潜层向量对齐的持续零样本学习算法

在持续学习多任务过程中,持续零样本学习旨在积累已见类知识,并用于识别未见类样本.然而,在连续学习过程中容易产生灾难性遗忘,因此,文中提出基于潜层向量对齐的持续零样本学习算法.基于交叉分布对齐变分自编码器网络框架,将当前任务与已学任务的视觉潜层向量对齐,增大不同任务潜层空间的相似性.同时,结合选择性再训练方法,提高当前任务模型对已学任务判别能力.针对不同任务,采用已见类视觉-隐向量和未见类语义-隐向量训练独立的分类器,实现零样本图像分类.在4个标准数据集上的实验表明文中算法能有效实现持续零样本识别任务,缓解算法的灾难性遗忘.

Continual Zero-Shot Learning Algorithm Based on Latent Vectors Alignment

Continual zero-shot learning aims to accumulate the knowledge of seen classes and utilize the knowledge for unseen classes recognition. However, catastrophic forgetting can easily occur in continual learning. Therefore, a continual zero-shot learning algorithm based on latent vectors alignment is proposed. Based on the cross and distribution aligned variational auto-encoder network, the visual latent vectors of current tasks and learned tasks are aligned to enhance the similarity of latent space of different tasks. Selective retraining is adopted to improve the discrimination ability of the current task model for learned tasks. For different tasks, the independent classifiers are trained with visual-hidden vectors of the seen classes and semantic-hidden vectors of the unseen classes to achieve zero-shot image classification. Extensive experiments on four standard datasets show that the proposed algorithm completes the continual zero-shot recognition task effectively and alleviates the catastrophic forgetting.