Enhancing Classification Accuracy on Limited Data via Unconditional GAN

Abstract

Despite significant advances in Deep Neural Networks (DNNs), these models often fall short in real-world scenarios, particularly when faced with a scarcity of training data. In this paper, we introduce a novel method that capitalizes on the power of Generative Adversarial Networks (GANs) to enhance performance in image classification tasks. Our approach specifically involves training the classifier by enforcing a consistency rule across generated unlabeled data synthesized from unconditional GANs. Through the implementation of our proposed methodology, we observed a substantial increase in accuracy - approximately 8.68% on the CIFAR-10 dataset compared to the baseline (which had an accuracy of 54.54%) trained with 500 real images. This notable enhancement in accuracy demonstrates the superiority of our method using class unconditional GANs over the previous techniques aiming to enhance accuracy using class Conditional GANs.