MetaZero: A Novel Meta-Learning Method Suitable for Generalized Zero-Shot Learning

Generalized zero-shot learning (GZSL) aims to recognize objects from both seen and unseen classes, when only the labeled examples from seen classes are provided. Most GZSL methods only optimize models based on seen classes but fail to explicitly mimic zero-shot learning settings that transfer knowledge from seen classes to unseen classes at the training time. Meta-learning has been introduced to mitigate this problem, but the models of current meta-learning approaches only learn to face the novel classes rather than learn to face the strong biased problem during the training, which is the main reason for the poor performance of most methods in GZSL. Humans can improve their knowledge with well-designed tests. According to this phenomenon, We propose a novel meta-learning method that is more suitable for GZSL termed MetaZero, aiming to alleviate the strong bias problem in GZSL. Specifically, MetaZero modifies the division of meta train set and meta test set, so that the model can directly face the strong bias problem in the meta test stage. In this way, with the iteration of training, the model can become an expert to solve the strong bias problem. Our extensive experiments manifest that our method achieves the significant gains than the SOTA GZSL methods.