基于低秩堆栈式语义自编码器的零样本学习

Zero-shot learning based on stacked semantic auto-encoder with low-rank embedding

在图像分类领域,现有的深度学习等方法在训练时需要大量有标注的数据样本,且无法识别在训练阶段未出现的类别。零样本学习能有效缓解此类问题。本研究基于堆栈式自编码器和低秩嵌入,提出了一种新的零样本学习方法,即基于低秩嵌入的堆栈语义自编码器(low-rank stacked semantic auto-encoder,LSSAE)。该模型基于编码-解码机制,编码器学习到一个具有低秩结构的投影函数,用于将图像的视觉特征空间、语义描述空间以及标签进行连接;解码阶段重建原始视觉特征。并通过低秩嵌入,使得学习到的模型在预见未见类别时能共享已见类的语义信息,从而更好地进行分类。本研究在五个常见的数据集上进行实验,结果表明LSSAE的性能优于已有的零样本学习方法,是一种有效的零样本学习方法。

In the field of image classification,existing methods such as deep learning require a large number of annotated samples for training and are unable to identify classes that do not appear in the training phase.Zero-shot learning tasks can effectively alleviate such problems.This study proposed a new zero-shot learning method,namely low-rank stacked semantic auto-encoder(LSSAE)based on stacked auto-encoder and low-rank embedding.The model was based on an encoding-decoding me-chanism where the encoder learned a projection function with a low-rank structure for concatenating the visual feature space,the semantic space and the labels.It reconstructed the original visual features in the decoding stage.And the low-rank embedding enabled the learned model to share the semantic information of the seen classes when anticipating the unseen classes for better classification.Experiments were conducted on five common datasets in this study,and the results show that the proposed LSSAE outperforms existing zero-shot learning methods which is an effective zero-shot learning method.