A novel individual-relational consistency for bad semi-supervised generative adversarial networks (IRC-BSGAN) in image classification and synthesis.

Semi-supervised learning leverages both labeled and unlabeled images for model training, addressing the scarcity of labeled data. However, challenges persist, including the determination of appropriate thresholds for pseudo-labeling, the effective utilization of uncertain unlabeled images, the absence of consistency regularization, and the oversight of inter-image relationships among images in low-density areas. This study introduces a novel approach named the Individual-Relational Consistency for Bad Semi-supervised Generative Adversarial Networks (IRC-BSGAN) to tackle these issues. IRC-BSGAN integrates bad adversarial training, consistency regularization, and pseudo-labeling to reduce error rates and enhance classifier performance. It includes various components, such as a bad generator network, a discriminator network, a classifier, and consistency regularization modules. IRC-BSGAN introduces new individual and relational consistency regularization losses on bad fake images in low-density areas, thereby generating informative images that precisely estimate the classifier's decision boundary. The proposed method ensures diversity and consistent labeling of bad fake images by integrating consistency mechanisms. It particularly focuses on low-density areas and extracts extra semantic details from these images by promoting local consistency and coherence among them. The effectiveness of IRC-BSGAN is realized by improving the pseudo-labeling of unlabeled images, especially for low-confidence unlabeled images. For the SVHN dataset with 1000 labeled training images and the CIFAR-10 dataset with 4000 labeled training images, the error rate reduced from 3.89 to 3.67 and from 7.29 to 6.17, respectively. Similarly, on the CINIC-10 dataset with 1000 labeled training images per class, IRC-BSGAN achieved a reduction in error rate from 19.38 to 15.45. On the COVID-19 dataset with 30 labeled training images, the error rate decreased from 7.41 to 5.55. [ABSTRACT FROM AUTHOR]