A novel dataset-specific feature extractor for zero-shot learning

Abstract

Most of the existing Zero-Shot Learning(ZSL) algorithms adopt pre-trained neural networks as their feature extractors. Since these pre-trained models are not specially designed for ZSL tasks, it is difficult to guarantee the stability and generalization ability of the ZSL algorithms due to the feature mismatch. To alleviate this problem, we propose a novel dataset-specific feature extractor for ZSL according to an attribute-based label tree. Specifically, an attribute-based label tree is firstly built via K-means clustering and then the information extracted from the label tree is used to fine-tune the parameters of the pre-trained models in order to make the extracted features more suitable for the current ZSL task. The experimental results on three typical ZSL datasets show that our approach can effectively improve the predictive accuracy of the existing ZSL algorithms and significantly accelerate their convergence rate. Additionally we explain the experimental phenomena from the perspective of feature visualization, which experimentally show that the features extracted by our method are much more separable than those of the original pre-trained models.