Adaptive Margin-based Contrastive Network for Generalized Zero-Shot Learning

Generalized zero-shot learning is a challenging problem that aims to recognize images from seen and unseen classes. Recent methods are costly and time-consuming or have a bias problem. To tackle this problem, we proposed an adaptive margin-based contrastive network that aims to distinguish similar classes in generalized zero-shot learning. The proposed method employs the architecture of transferable contrastive network to classify unseen classes and adaptive margin to transfer discriminative knowledge. Experiments on the AwA2 dataset demonstrate competitive results against state-of-the-art benchmarks.