Regularized label relaxation-based stacked autoencoder for zero-shot learning

Recently, Zero-Shot Learning (ZSL) has gained great attention due to its significant classification performance for novel unobserved classes. As seen and unseen classes are completely disjoint, the current ZSL methods inevitably suffer from the domain shift problem when transferring the knowledge between the observed and unseen classes. Additionally, most ZSL methods especially those targeting the semantic space may cause the hubness problem due to their use of nearest-neighbor classifiers in high-dimensional space. To tackle these issues, we propose a novel pathway termed Regularized Label Relaxation-based Stacked Autoencoder (RLRSA) to diminish the domain difference between seen and unseen classes by exploiting an effective label space, which has some notable advantages. First, the proposed method establishes the tight relations among the visual representation, semantic information and label space using via the stacked autoencoder, which is beneficial for avoiding the projection domain shift. Second, by incorporating a slack variable matrix into the label space, our RLRSA method has more freedom to fit the test samples whether they come from the observed or unseen classes, resulting in a very robust and discriminative projection. Third, we construct a manifold regularization based on a class compactness graph to further reduce the domain gap between the seen and unseen classes. Finally, the learned projection is utilized to predict the class label of the target sample, thus the hubness issue can be prevented. Extensive experiments conducted on benchmark datasets clearly show that our RLRSA method produces new state-of-the-art results under two standard ZSL settings. For example, the RLRSA obtains the highest average accuracy of 67.82% on five benchmark datasets under the pure ZSL setting. For the generalized ZSL task, the proposed RLRSA is still highly effective, e.g., it achieves the best H result of 58.9% on the AwA2 dataset.