Attract or Distract Exploit the Margin of Open Set论文总结

Attract or Distract Exploit the Margin of Open Set论文总结

该论文研究的问题

该论文研究的是Open set domain adaptation问题,其中源域中没有未知类,而作者将目标域中的未知样本统称为一类:unknown class。文章的目标是在没有目标域标签的情况下,同时处理域转移未知对象的识别(这里的识别只是从目标域中的已知类中区分推离,而不是对未知类进行类别的详细识别)。

具体思路

扩大两种边界:

边界一:已知类之间的边界,目标是使已知的类更加可分

​ 1.源域与目标域之间的已知类:

​ 使用语义类别对齐semantic categorical alignment (SCA)),通过将目标域已知类与源域的已知类进行匹配对齐,间接提高二者的可分性。

​ 2.源域之间的已知类:

​ 已知类之间的差别可以使用交叉熵所示来实现,但是作者使用了对比中心损失(contrastive-center loss)来进一步增强类别之间的可分离性。

边界二:未知样本和已知类之间的边界,期望将未知的类推离决策边界

​ 使用了**语义对比映射( semantic contrastive mapping (SCM) )**将未知类推离边界:设计对比损失,使未知类和已知类之间的差值大于已知类之间的差值

由于目标标签不可用,作者使用每次迭代的网络预测作为目标标签的假设来执行语义分类对齐和语义对比映射。

论文创新点

我认为文章根本的创新点在于将源域与目标域的样本看作类似点集,通过包裹在SCA、SCM模型中的各种损失函数的设定、类别中心的计算与迭代更新将抽象问题转化为求解类别中心、同一类的边界聚拢、不同类的边界排斥等具体且易理解的操作。

三大模型作用

1.预训练对抗网络中的生成器与鉴别器,先通过Adversarial Domain Adaptation初步进行源域各个类别的区分、目标域与源域相应类别的中心对齐、并初步形成目标域中已知样本与未知样本的边界

2.通过Semantic Categorical Alignment,使每个已知类更加集中,源和目标之间的对齐更加准确(通过分类中心对齐损失实现),拉近已知类内部的表示,而加大各个已知类之间的区分(通过对比中心损失实现)

3.使用Semantic Contrastive Mapping,对于目标域中的非中心样本,(使用对比损失函数)鼓励已知样本向其中心靠近,并强制未知样本远离所有已知类的中心

  • 0
    点赞
  • 0
    收藏
    觉得还不错? 一键收藏
  • 0
    评论
Publishing your research in an international journal is key to your success in academia. This guide is based on a study of referees' reports and letters from journal editors on reasons why papers written by non-native researchers are rejected due to problems with English usage. It draws on English-related errors from around 5000 papers written by non-native authors, 500 abstracts by PhD students, and over 1000 hours of teaching researchers how to write and present research papers. With easy-to-follow rules and tips, and with examples taken from published and unpublished papers, you will learn how to: prepare and structure a manuscript increase readability and reduce the number of mistakes you make in English by writing concisely, with no redundancy and no ambiguityplan and organize your paper, and structure each paragraph and each sentence so that the reader can easily follow the logical build-up towards various conclusionswrite a title and an abstract that will attract attention and be readdecide what to include in the various parts of the paper (Introduction, Methodology, Discussion etc) select from over 700 useful phrases highlight your claims and contributionavoid plagiarism and make it 100% clear whether you are referring to your own work or someone else’schoose the correct tenses and style (active or passive) Other books in the series: English for Presentations at International Conferences English for Academic Correspondence and Socializing English for Research: Usage, Style, and Grammar English for Academic Research: Grammar / Vocabulary / Writing Exercises Adrian Wallwork is the author of more than 20 ELT and EAP textbooks. He has trained several thousand PhD students and academics from 35 countries to prepare and give presentations. Since 1984 he has been revising research papers, and in 2009 he set up englishforacademics.com – a proofreading and editing service specifically for researchers.
WIDE bandgap devices, such as silicon carbide (SiC) metal–oxide–semiconductor field-effect transis- tors (MOSFETs) present superior performance compared to their silicon counterparts [1]. Their lower ON-state resistance and faster switching capability attract lots of interest in high-power- density applications [2]. Faster switching speed enables lower switching loss and higher switching frequency, which is benefi- cial to high-efficiency and high power density. However, severe electromagnetic interference (EMI) and transient overvoltage issues caused by fast switching speed jeopardize the power quality and reliability of converters [3], [4]. Therefore, there is a tradeoff between efficiency and reliability in the choice of switching speed. An optimized design should ensure theoperation within both safe-operation-area and EMI limits, and switching loss should be as small as possible. A prediction method of switching performance is important and helpful for designer to evaluate and optimize converter design. The most concerned switching characteristics are switching loss, dv/dt, di/dt, and turn-ON/OFF overvoltage generally. These characteristics are crucial for the design of heatsink, filter, and gate driver. Related discussions have been presented in many existing research articles as following.请将这一段进行以下要求,Move analysis 语步(内容成分)分析; Language devices和实现该功能的语言手段(某些关键专有名词提供汉语翻译)
06-13

“相关推荐”对你有帮助么?

  • 非常没帮助
  • 没帮助
  • 一般
  • 有帮助
  • 非常有帮助
提交
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值