• 主题
高级搜索  >
历史搜索 清空历史
首页> 外文会议>International Conference on Control, Automation and Information Sciences >Parasitic Network: Learning-Based Network Downsizing of Very Deep Neural Networks for Computer Vision
【24h】

Parasitic Network: Learning-Based Network Downsizing of Very Deep Neural Networks for Computer Vision

机译：寄生网络：用于计算机视觉的非常深度神经网络的基于学习的网络精简

获取原文
页面导航
  • 摘要
  • 著录项
  • 相似文献
  • 相关主题

摘要

In recent research on deep neural network (DNN), network downsizing is one of practical issues for computational and memory efficiency. Specifically, downsizing or compression of networks minimizing the performance degradation becomes a critical problem for deployment of DNNs on resource-limited environments such mobile or embedded platforms. In this paper, we propose a compressed network called the parasitic network (PN) inspired by the relationship between a parasite and host in nature. The concept of the parasitic network is straightforward. The host network provides their mapping results in each layer to the PN as a feed. The PN that much shallower than the host network is trained based on given information from the host network. We demonstrate efficiency of our approach to the network downsizing in image classification and object detection problems which have conquered by the deeper and bigger networks. The experimental results show that PN can provide sample performance to their host network even though their architectural scale is much smaller.
机译：在最近的深度神经网络（DNN）研究中，网络精简是计算和存储效率方面的实际问题之一。具体而言，缩小网络规模或压缩网络以将性能降低降至最低成为将DNN部署在资源受限的环境（如移动平台或嵌入式平台）上的关键问题。在本文中，我们提出了一种受寄生虫和宿主自然关系启发的压缩网络，称为寄生网络（PN）。寄生网络的概念很简单。主机网络将它们在每层中的映射结果作为提要提供给PN。根据来自主机网络的给定信息，对比主机网络浅得多的PN进行训练。我们展示了在缩小图像分类和目标检测问题中缩小网络规模的方法的效率，这些问题已被更深和更大的网络所征服。实验结果表明，即使PN的体系结构规模小得多，PN也可以为其主机网络提供示例性能。

著录项

相似文献

  • 外文文献
  • 中文文献
  • 专利
获取原文

客服邮箱：kefu@zhangqiaokeyan.com

京公网安备：11010802029741号 ICP备案号：京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司©版权所有

  • 客服微信

  • 服务号