人工智能培训

搜索

人工智能教程:IGCV3:高效深层神经网络的交错低秩群卷积(IGCV3: Interleaved Low-Rank Group Convolutions for

[复制链接]
ppkosk168 发表于 2018-6-4 08:28:12 | 显示全部楼层 |阅读模式
ppkosk168 2018-6-4 08:28:12 1061 0 显示全部楼层
人工智能教程:IGCV3:高效深层神经网络的交错低秩群卷积(IGCV3: Interleaved Low-Rank Group Convolutions for Efficient Deep Neural  Networks)在本文中,我们有兴趣构建轻量级和高效的卷积神经网络。受两种设计模式的成功启发,如结构化稀疏内核的组合,例如交错组卷积(IGC)和低秩内核的组合(例如瓶颈模块),我们研究了这两种设计模式的组合,使用组成稀疏的低阶内核,形成卷积核。我们并没有在信道上引入补充条件,而是引入了一种宽松的补充条件,该条件是通过强加超信道的互补条件来制定的,以指导产生密集卷积核的设计。最终的网络被称为IGCV3。我们凭经验证明,低级别和稀疏内核的组合提高了我们提议的方法对状态,IGCV2和MobileNetV2的性能和优势,优于CIFAR和ImageNet上的图像分类以及COCO上的对象检测。
In this paper, we are interested in building lightweight and efficientconvolutional neural networks.Inspired by the success of two design patterns,composition of structured sparse kernels, eg, interleaved group convolutions(IGC), and composition of low-rank kernels, eg, bottle-neck modules, we studythe combination of such two design patterns, using thecomposition ofstructured sparse low-rank kernels, to form a convolutional kernel.Rather thanintroducing a complementary condition over channels, we introduce a loosecomplementary condition, which is formulated by imposing the complementarycondition over super-channels, to guide the design for generating a denseconvolutional kernel.The resulting network is called IGCV3.We empiricallydemonstrate that the combination of low-rank and sparse kernels boosts theperformance and the superiority of our proposed approach to thestate-of-the-arts, IGCV2 and MobileNetV2 over image classification on CIFAR andImageNet and object detection on COCO.人工智能教程:IGCV3:高效深层神经网络的交错低秩群卷积(IGCV3: Interleaved Low-Rank Group Convolutions for Efficient Deep Neural  Networks) Okk1VyEF8ZFyene8.jpg
URL地址:https://arxiv.org/abs/1806.00178     ----pdf下载地址:http://arxiv.org/pdf/1806.00178    ----人工智能教程:IGCV3:高效深层神经网络的交错低秩群卷积(IGCV3: Interleaved Low-Rank Group Convolutions for Efficient Deep Neural  Networks)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

ppkosk168当前离线
新手上路

查看:1061 | 回复:0

快速回复 返回顶部 返回列表