Discriminative Latent Attribute Autoencoder for Zero-Shot Learning

Zero-shot learning aims to recognize classes without labeled samples training. One of the most widely utilized methodologies for ZSL is to learn a shared semantic embedding space with attributes directly. However, the user-defined attributes are not necessarily discriminative but treated directly and independently. Meanwhile, most approaches have the domain shift problem that means classes domain (seen and unseen) are different. To address these two problem, the paper presents a new model for ZSL addressing these two issues. Firstly, our model constructs a latent attribute space with the attribute space and the similarity space to make the latent attributes semantic and discriminative. Secondly, we apply two reconstruction constraints to the feature space and the similarity space with autoencoder. Thirdly, we combine the latent attribute space and the similarity space for ZSL prediction. We evaluate our model performance on two benchmark datasets, which is competitive to the existed approaches.