Generalized Zero Shot Learning

Generalized zero-shot learning (GZSL) aims to train models that can classify objects from categories unseen during training, leveraging semantic information (e.g., attributes, textual descriptions) to bridge the gap between seen and unseen classes. Current research focuses on mitigating bias towards seen classes, often employing generative models (e.g., diffusion models, VAEs) to synthesize unseen data, or leveraging the capabilities of large language models and vision-language models (like CLIP) to improve semantic alignment and feature extraction. These advancements are significant because GZSL addresses the limitations of traditional supervised learning, enabling more robust and adaptable AI systems for applications where labeled data for all classes is scarce or unavailable.