Scholarlink

<p>The primary aim of this study is to explore how deep learning, enhanced by GPU optimization, can improve the accuracy and efficiency of singlecell protein localization classification using the Human Protein Atlas (HPA) dataset. This study utilized Visual Geometry Group 26 (VGG16) and the InceptionV3 deep learning models with the Human Protein Atlas dataset for protein subcellular localization. To handle computational demands and manual labeling limitations, GPU optimization techniques, including mixed precision training, batch size optimization, memory pinning, and model pruning, were integrated. These optimizations aim to enhance training efficiency by reducing average time per epoch and total training time. Models with more data are more precise, while those with minimal data, such as for disease Rods and Rings, have a precision of 0. Parallelization techniques improve training efficiency, especially with GPU utilization. Mixed precision training reduces training times, particularly for InceptionV3. Batch size optimization and memory pinning also contribute to time reductions, with varying impact across models. Model pruning effectively reduces average training time, highlighting the potential for computational efficiency gains through architectural optimizations. This study emphasizes the importance of GPU optimization techniques (mixed precision training, batch size optimization, memory pinning, and model pruning) in improving the efficiency of deep learning workflows for protein subcellular localization classification. The findings suggest that InceptionV3 generally benefits more than VGG16 from these techniques, advocating for their customized use based on specific deep learning architectures to enhance computational efficiency and reduce training times in computational biology research.</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

GPU-Optimized Deep Learning Methods for Single-Cell Protein Localization Classification: A Comparative Analysis

  • 期刊论文

  • 发表日期：

    2025年6月10日

查看全文
  • Anushka Vijaykumar Naik,
  • Nidhi Patel,
  • Aagam Bakliwal,
  • Prajyot Kale,
  • Amit D. Joshi,
  • Suraj T. Sawant

摘要

The primary aim of this study is to explore how deep learning, enhanced by GPU optimization, can improve the accuracy and efficiency of singlecell protein localization classification using the Human Protein Atlas (HPA) dataset. This study utilized Visual Geometry Group 26 (VGG16) and the InceptionV3 deep learning models with the Human Protein Atlas dataset for protein subcellular localization. To handle computational demands and manual labeling limitations, GPU optimization techniques, including mixed precision training, batch size optimization, memory pinning, and model pruning, were integrated. These optimizations aim to enhance training efficiency by reducing average time per epoch and total training time. Models with more data are more precise, while those with minimal data, such as for disease Rods and Rings, have a precision of 0. Parallelization techniques improve training efficiency, especially with GPU utilization. Mixed precision training reduces training times, particularly for InceptionV3. Batch size optimization and memory pinning also contribute to time reductions, with varying impact across models. Model pruning effectively reduces average training time, highlighting the potential for computational efficiency gains through architectural optimizations. This study emphasizes the importance of GPU optimization techniques (mixed precision training, batch size optimization, memory pinning, and model pruning) in improving the efficiency of deep learning workflows for protein subcellular localization classification. The findings suggest that InceptionV3 generally benefits more than VGG16 from these techniques, advocating for their customized use based on specific deep learning architectures to enhance computational efficiency and reduce training times in computational biology research.