Deconstructed Generation-Based Zero-Shot Model

Recent research on Generalized Zero-Shot Learning (GZSL) has focused primarily generation-based methods. However, current literature overlooked the fundamental principles of these methods and made limited progress in a complex manner. In this paper, we aim to deconstruct generator-classifier framework provide guidance for its improvement extension. We begin by breaking down generator-learned unseen class distribution into class-level instance-level distributions. Through our analysis role two types distributions solving GZSL problem, generalize focus approach, emphasizing importance (i) attribute generalization generator learning (ii) independent classifier with partially biased data. present simple method based that outperforms SotAs four public datasets, demonstrating validity deconstruction. Furthermore, proposed remains effective even without generative model, representing step towards simplifying structure. Our code is available at https://github.com/cdb342/DGZ.