Graph active learning for GCN-based zero-shot classification

Abstract

In the supervised learning approach, classification models can only categorize objects into seen classes for which labeled data instances are available for training. Zero-shot learning, especially the recent graph neural network-based zero-shot learning, is commonly accepted as an effective approach to address this limitation by exploiting relations between seen classes and unseen classes. However, lots of seen classes are still necessary for most zero-shot learning frameworks to infer the classification model for unseen classes, especially in the image classification task. In this paper, we propose an active learning framework of graph convolutional network (GCN)-based zero-shot learning for image classification and design a new active learning algorithm called GAZL that can enable the zero-shot learning model to achieve a higher performance with a fixed amount of seen classes. Our algorithm extends the k-center algorithm with a new Laplacian energy-based strategy to select the most representative and crucial classes as seen classes for training. Extensive experiments on ImageNet demonstrate that our active learning method is superior to a wide spectrum of active learning methods for GCN-based zero-shot learning.