Scholarlink

<p>In this paper, we introduce ROS-GPU, a novel large-scale image classification algorithm that replaces the conventional multi-layer perceptron (MLP) classification head with an ensemble of Random Oblique Stumps (ROS) trained using a GPU-optimized Multi-Class Linear Discriminant Analysis (MC-LDA). Unlike prior ROS-based approaches, the proposed algorithm unifies oblique stump construction and multi-class discrimination through a fully parallel One-Versus-All (OVA) MC-LDA formulation, specifically designed for efficient execution on modern GPU architectures. To address class imbalance in large-scale multi-class datasets, the algorithm incorporates an under-sampling strategy within the OVA scheme to improve class balance and stabilize scalable stump learning. The proposed ROS-GPU algorithm is implemented using CUDA and cuBLAS and is evaluated on the ImageNet benchmark. Experimental results demonstrate that ROS-GPU significantly reduces training time while achieving competitive classification accuracy compared to established learning methods. In particular, ROS-GPU completes ImageNet training in 5.86&#xa0;min while attaining an accuracy of 89.44%. These results show the effectiveness of combining random oblique ensembles with GPU-parallel discriminant learning, offering a computationally efficient and scalable alternative to traditional deep classification heads for high-dimensional image recognition tasks.</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

Classifying the ImageNet dataset with ROS-GPU

  • 期刊论文

  • 发表日期：

    2026年6月14日

查看全文
  • Thanh-Nghi Do,
  • Hoai An Le Thi

摘要

In this paper, we introduce ROS-GPU, a novel large-scale image classification algorithm that replaces the conventional multi-layer perceptron (MLP) classification head with an ensemble of Random Oblique Stumps (ROS) trained using a GPU-optimized Multi-Class Linear Discriminant Analysis (MC-LDA). Unlike prior ROS-based approaches, the proposed algorithm unifies oblique stump construction and multi-class discrimination through a fully parallel One-Versus-All (OVA) MC-LDA formulation, specifically designed for efficient execution on modern GPU architectures. To address class imbalance in large-scale multi-class datasets, the algorithm incorporates an under-sampling strategy within the OVA scheme to improve class balance and stabilize scalable stump learning. The proposed ROS-GPU algorithm is implemented using CUDA and cuBLAS and is evaluated on the ImageNet benchmark. Experimental results demonstrate that ROS-GPU significantly reduces training time while achieving competitive classification accuracy compared to established learning methods. In particular, ROS-GPU completes ImageNet training in 5.86 min while attaining an accuracy of 89.44%. These results show the effectiveness of combining random oblique ensembles with GPU-parallel discriminant learning, offering a computationally efficient and scalable alternative to traditional deep classification heads for high-dimensional image recognition tasks.