Abstract
Generalized zero-shot learning(GZSL) aims to classify samples from seen andunseen labels, assuming unseen labels are not accessible during training.Recent advancements in GZSL have been expedited by incorporatingcontrastive-learning-based (instance-based) embedding in generative networksand leveraging the semantic relationship between data points. However, existingembedding architectures suffer from two limitations: (1) limiteddiscriminability of synthetic features' embedding without consideringfine-grained cluster structures; (2) inflexible optimization due to restrictedscaling mechanisms on existing contrastive embedding networks, leading tooverlapped representations in the embedding space. To enhance the quality ofrepresentations in the embedding space, as mentioned in (1), we propose amargin-based prototypical contrastive learning embedding network that reaps thebenefits of prototype-data (cluster quality enhancement) and implicit data-data(fine-grained representations) interaction while providing substantial clustersupervision to the embedding network and the generator. To tackle (2), wepropose an instance adaptive contrastive loss that leads to generalizedrepresentations for unseen labels with increased inter-class margin. Throughcomprehensive experimental evaluation, we show that our method can outperformthe current state-of-the-art on three benchmark datasets. Our approach alsoconsistently achieves the best unseen performance in the GZSL setting.