Deconstructed Generation-Based Zero-Shot Model

Generation-based methods have captured most of the recent attention in Zero-Shot Learning research. In this paper, we attempt to deconstruct the generator-classifier framework to guide its improvement and extension. We begin by analyzing the generator-learned instance-level distribution by alternating it with a Gaussian distribution. Then we reveal the roles of the class-level distribution and the instance-level distribution learned by the generator in classifier training by decomposing the classifier gradients. We finally conclude with the guidelines for improving the generator-classifier framework from the deconstruction of the generator and the classifier, i.e., (i) The key for the ZSL generator is attribute generalization; and (ii) classifier learning emphasizes mitigating the impact of pseudo unseen samples on decision boundaries between seen classes during training, and reducing the seen-unseen bias. We propose a simple method based on the guidelines. Without complex designs, the proposed method outperforms the state of the art on four public ZSL datasets, which demonstrates the validity of the proposed guidelines. The proposed method is still effective when replacing the generative model with an attribute-to-visual center single mapping model, demonstrating its strong transferability. Codes will be public upon acceptance.