Scholarlink

<p>Contrastive learning has arisen as a potent method in self-supervised learning (SSL), seeking to acquire rich and significant representations from unlabeled data by differentiating between positive and negative samples. However, the efficacy of contrastive learning is heavily contingent upon many aspects, including feature representation, hyperparameters, and data augmentation techniques, which are often manually tuned and suboptimal. To address these challenges, this research presents a unique framework named HOCLGA that uses Genetic Algorithms (GAs) to automate and optimize these critical components for contrastive learning. In fact, by combining GA with contrastive learning, we provide a systematic method for optimizing feature representations, improving the quality of learned representations, adjusting hyperparameters, investigating data augmentation techniques, and formulating SSL challenges. This study examines the use of self-supervised learning to enhance the precision of identifying damaged regions in medical imaging using High-Intensity Focused Ultrasound (HIFU). The framework is pretrained on large-scale datasets (ImageNet, DeepLesion, Breast Ultrasound) and fine-tuned for ultrasound image analysis. Experimental findings indicate that the proposed method may improve representation learning efficacy and provide substantial progress in medical picture interpretation. Specifically, it achieves the highest accuracy of 94.3% on the Deep Lesion (CT image) dataset and the highest precision of 96.7% on the Breast Ultrasound dataset. Key innovations include a custom contrastive loss function leveraging Cauchy distributions for similarity scoring and a GA-driven optimization of ResNet architectures, augmentation policies, and hyperparameters.</p><p>Our findings highlight HOCLGA’s ability to improve representation quality and diagnostic accuracy in medical imaging, particularly for scenarios with limited labeled data. The framework’s evolutionary approach offers a scalable solution for SSL optimization, with potential applications beyond healthcare. This work advances the integration of metaheuristics and SSL, providing a robust tool for automated, high-precision medical image analysis. </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

Evolutionary optimization for enhanced self-supervised learning: leveraging genetic algorithms for representation learning

  • 期刊论文

  • 发表日期：

    2025年6月25日

查看全文
  • Matineh Zavar,
  • Hamid Reza Ghaffari,
  • Hamid Tabatabaee

摘要

Contrastive learning has arisen as a potent method in self-supervised learning (SSL), seeking to acquire rich and significant representations from unlabeled data by differentiating between positive and negative samples. However, the efficacy of contrastive learning is heavily contingent upon many aspects, including feature representation, hyperparameters, and data augmentation techniques, which are often manually tuned and suboptimal. To address these challenges, this research presents a unique framework named HOCLGA that uses Genetic Algorithms (GAs) to automate and optimize these critical components for contrastive learning. In fact, by combining GA with contrastive learning, we provide a systematic method for optimizing feature representations, improving the quality of learned representations, adjusting hyperparameters, investigating data augmentation techniques, and formulating SSL challenges. This study examines the use of self-supervised learning to enhance the precision of identifying damaged regions in medical imaging using High-Intensity Focused Ultrasound (HIFU). The framework is pretrained on large-scale datasets (ImageNet, DeepLesion, Breast Ultrasound) and fine-tuned for ultrasound image analysis. Experimental findings indicate that the proposed method may improve representation learning efficacy and provide substantial progress in medical picture interpretation. Specifically, it achieves the highest accuracy of 94.3% on the Deep Lesion (CT image) dataset and the highest precision of 96.7% on the Breast Ultrasound dataset. Key innovations include a custom contrastive loss function leveraging Cauchy distributions for similarity scoring and a GA-driven optimization of ResNet architectures, augmentation policies, and hyperparameters.

Our findings highlight HOCLGA’s ability to improve representation quality and diagnostic accuracy in medical imaging, particularly for scenarios with limited labeled data. The framework’s evolutionary approach offers a scalable solution for SSL optimization, with potential applications beyond healthcare. This work advances the integration of metaheuristics and SSL, providing a robust tool for automated, high-precision medical image analysis.