Class-wise and instance-wise contrastive learning for zero-shot learning based on VAEGAN

B. Zheng

doi:10.1016/j.eswa.2025.126671

Return

Class-wise and instance-wise contrastive learning for zero-shot learning based on VAEGAN

2025-05-01

PRE

B. Zheng

Zhanshan Li

李静瑶 (Jingyao Li) *

DOI：10.1016/j.eswa.2025.126671

Original

Original request for help

Save

Abstract

En 中文

Generalized zero-shot learning (GZSL) is a further extension of the zero-shot learning field, aiming to address the problem of simultaneous classification of known and unknown categories within a unified framework. Variational Autoencoder Generative Adversarial Network (VAEGAN) models are widely favored for generating high-quality, diverse features. Additionally, contrastive learning in the realm of zero-shot learning enables effective utilization of unlabeled data by learning the similarity and dissimilarity between samples without requiring sample labels. This paper proposes a contrastive learning framework based on VAEGAN for class-wise and instance-wise learning. Firstly, we introduce a feature feedback module to facilitate inter-class contrast by further processing the features optimized through inter-class contrast, along with visually processed features, and then feed them back to the decoder and encoder within the VAEGAN model to maintain consistency between data and generated features. Secondly, instance-level contrastive learning is introduced, enabling the model to perform instance-level contrast alongside inter-class contrast. We embed the generated features into an embedding space and apply nonlinear transformations, allowing instances of the same class to approach each other while instances of different classes move away, further enhancing classification performance. Finally, we validate our method on four benchmark datasets of GZSL, demonstrating its superiority over state-of-the-art methods.

Keywords:

Zero-shot learning

Machine learning

Generative models

Contrastive learning

Image classification