Scholarlink

<p>Semi-supervised learning has emerged as a powerful approach to addressing challenges posed by limited labeled data, particularly in image classification tasks. This study investigates the integration of wavelet transforms, vision transformers, dropout layers, interpolation regularization, and ensemble within bad generative adversarial networks (GANs) to enhance model performance. The primary research question addresses how these components can effectively manage pseudo-label fluctuations and improve classification accuracy using unlabeled images. Previous methods have struggled with low-confidence pseudo-labels, leading to noisy predictions. Our framework aims to fill this gap by leveraging a novel loss function and a unique discriminator architecture. The proposed method, based on the wavelet-transformer discriminator, yielded an error rate of 2.84 ± 0.15 on SVHN, 5.79 ± 0.08 on CIFAR-10, 21.06 ± 0.09 on STL-10, and 15.20 ± 0.08 on CINIC-10 dataset, using 1000, 4000, 5000, and 10,000 labeled images, respectively. These findings underscore the framework’s potential to enhance classification in semi-supervised learning scenarios, contributing to more accurate models in real-world applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A novel wavelet-transformer discriminator for semi-supervised GANs with controlled regularization and ensemble techniques

  • 期刊论文

  • 发表日期：

    2025年9月1日

查看全文
  • Mohammad Saber Iraji

摘要

Semi-supervised learning has emerged as a powerful approach to addressing challenges posed by limited labeled data, particularly in image classification tasks. This study investigates the integration of wavelet transforms, vision transformers, dropout layers, interpolation regularization, and ensemble within bad generative adversarial networks (GANs) to enhance model performance. The primary research question addresses how these components can effectively manage pseudo-label fluctuations and improve classification accuracy using unlabeled images. Previous methods have struggled with low-confidence pseudo-labels, leading to noisy predictions. Our framework aims to fill this gap by leveraging a novel loss function and a unique discriminator architecture. The proposed method, based on the wavelet-transformer discriminator, yielded an error rate of 2.84 ± 0.15 on SVHN, 5.79 ± 0.08 on CIFAR-10, 21.06 ± 0.09 on STL-10, and 15.20 ± 0.08 on CINIC-10 dataset, using 1000, 4000, 5000, and 10,000 labeled images, respectively. These findings underscore the framework’s potential to enhance classification in semi-supervised learning scenarios, contributing to more accurate models in real-world applications.