Decentralized Contrastive Learning for generalized zero-shot image classification

Abstract

Generalized zero-shot learning (GZSL) aims to learn a model on known classes that can adapt to a test set comprising both known and unknown classes. Recent GZSL research in image classification has made significant progress by utilizing representation learning techniques. However, the challenge of generating discriminative representations for fine-grained classes with close relevance remains unresolved. To address this problem, we introduce a Decentralized Contrastive Learning (DCL) framework that seamlessly integrates a nested Wasserstein GAN (WGAN) with decentralized contrastive representation learning. Our nested WGAN incorporates the representation learning module within the discriminator, enabling the model to simultaneously train the representations and differentiate them in a synergistic manner. Moreover, our decentralized contrastive learning module enhances the discriminative nature of representations by preserving calibration based on class information without additional parameters during training. We further provide theoretical analysis for DCL, uncovering its superiority in learning discriminative representations and its robustness in handling mixed features. Experiments on show that DCL outperforms the state-of-the-art models by margins of about 3%, 4% and 3% on CUB, SUN and aPY datasets.