Learning by Asking Questions

最新推荐文章于 2020-06-24 10:56:49 发布
原创 最新推荐文章于 2020-06-24 10:56:49 发布 · 507 阅读 · 1 ·
CC 4.0 BY-SA版权
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
文章标签：

#Paper

本文介绍了一种名为学习通过询问(LBA)的交互式学习框架，该框架应用于视觉问题回答(VQA)任务中，旨在使智能视觉系统更加自然地学习和发展。与传统的VQA相比，LBA更接近于人类的学习方式，能够自动发现从易到难的学习路径，并生成更多有效的训练数据。

论文地址：http://openaccess.thecvf.com/content_cvpr_2018/papers/Misra_Learning_by_Asking_CVPR_2018_paper.pdf

摘要

       我们介绍一个交互式的学习框架来发展和测试智能视觉系统，叫做学习通过询问（LBA）。我们挖掘LBA在内容上，关于视觉问题回答（VQA）任务。LBA不同于标准的VQA训练，在很多问题上，其中不被注视在训练过程中，学习者必须询问问题，它想回答的。因此，LBA比较贴近模仿自然学习和有潜在的较多有效数据比传统VQA集。我们展示了一个模型，能够表现LBA在CLEVR数据集上，以及表明它是自动发现一个由易到难课程，当学习交互从一个对话中。我们的LBA生成数据和有较多有效样本。我们也展示我们的模型问问题，普遍是先进的比VQA模型，新颖的在测试时间分配上。

LBA范例

                           

LBA系统架构

不同迭代次数下

了解更多关于《计算机视觉与图形学》相关知识，请关注公众号：

下载我们视频中代码和相关讲义，请在公众号回复：计算机视觉课程资料

【论文阅读】Learning to Retrieve Reasoning Paths over Wikipedia Graph for Question Answering
没有胡子的猫Asimov
07-10 512
Learning to Retrieve Reasoning Paths over Wikipedia Graph for Question Answering 论文：https://arxiv.org/abs/1911.10470 代码：https://github.com/AkariAsai/learning_to_retrieve_reasoning_paths 学习在维基百科中检索问题的推理路径 基于推理路径 任务 从维基百科中提取推理路径实现多轮问答。 多轮问答： 需要结合多篇文档的“
Learning Prompt
清风冰竹
01-18 1378
说明：这是我的学习笔记，很多内容转自网络，请查阅文章末尾的参考资料。Character ThisistheroleyouwantChatGPTtoplay.Itcouldbeanexpert,acelebrity,anartistoranidiot.Makeitclearwithanintroductorystatementlike:Request ThisisthetaskyouwantChatGPTtodoforyou.You
Learning by Asking Questions 导读
JayShaun的博客
01-11 701
论文链接Arxiv 本文所提出的LBA不同于标准的VQA模型，LBA在训练时大部分question是不给出的，学习器必须问出它想回答的问题，这样做可以达到与环境交互的目的，就像是课堂上一个聪明的学生会问出有用的好问题来促进自己学习。大体步骤如下： 1）学习器需要一个语言模型来生成问题； 2）学习器需要理解视觉图像以确保问出的问题与图像是相关的； 3）学习器应该有“自知之明”，能够自我评价，换句话说...
总结 CVPR 2018 显著性目标检测文章
taxuewuhenxiaoer
01-09 590
总结 CVPR 2018 显著性目标检测文章 Progressive Attention Guided Recurrent Network for Salient Object Detection Learning to Promote Saliency Detectors Salient Object Detection Driven by Fixation Prediction A Bi-D...
Learning to Generate Questions by Learning What not to Generate阅读笔记
xulijun_811的博客
06-24 1290
再次调试代码后 简单记录了项目的结构和主要流程 一 加载数据集，用spacy进行预处理，提取词汇特征，保存成.pkl文件。 二 做嵌入，通过glove文件嵌入初始化每个向量，原文是glove.500B.100d.txt,运行内存不足，我换成了glove.6B.50d.txt.在prepro_util.py文件里做单词嵌入、答案位置嵌入、ner、pos、dep、iob等嵌入，得到嵌入矩阵和字典，构建特征实例。生成output里txt文件。 三 线索预测器 对于每个训练示例，我们都会得到输入通道和输.
深度学习Deep learning
GarfieldEr007的专栏
03-20 4684
In the last chapter we learned that deep neural networks are often much harder to train than shallow neural networks. That's unfortunate, since we have good reason to believe that if we could train
Quotes on Learning
CodingIsFun
08-01 1732
A good teacher can inspire hope, ignite the imagination, and instill a love of learning. Brad Henry There is no end to education. It is not that you read a book, pass an examination, and finish with...
Awesome Machine Learning
a1424262219的博客
08-22 681
Awesome Machine Learning A curated list of awesome machine learning frameworks, libraries and software (by language). Inspired by awesome-php. If you want to contribute to this ...
Learning Boost C++ Libraries 原版pdf by Mukherjee
04-27
In using many more Boost libraries to write software since then, I often found myself digging deep into the documentation, or asking questions on the mailing list and online forums to understand ...
Machine Learning for OpenCV
08-15
OpenCV machine learning connects the fundamental theoretical principles behind machine learning to their practical applications in a way that focuses on asking and answering the right questions....
GA-Net: Guided Aggregation Net for End-to-end Stereo Matching
CSS360的博客
06-26 3907
Code：https://github.com/feihuzhang/GANet 摘要 在立体匹配任务中，匹配损失配准是至关重要的在传统方法和深度神经网络模型中，为了精确估计差异。我们提出了两个新颖的神经网络层，目的去捕获一个局部和全局图像的成本依赖性。第一个是半全局聚集层，它是半监督匹配的一个可微近似，第二个是局部引导聚集层，它是使用了一个传统的损失滤波策略来简化结构。 ...
An Attention Enhanced Graph Convolutional LSTM Network for Skeleton-Based Action Recognition
CSS360的博客
06-30 2898
论文链接：https://arxiv.org/pdf/1902.09130.pdf 摘要 基于骨架的动作识别是一个非常重要的任务，需要足够的理解移动特点一个人活动，从给定的骨架序列。最近的研究已经展示，探索空间和时间的特征在骨架序列上，是非常重要的对于这个任务。尽管，如何去有效去提取有判断力的空间和时间特征仍然是一个具有挑战的问题。在这篇文章中，我们提出一个新颖的注意增强图卷积L...
Pseudo-LiDAR from Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Dr
CSS360的博客
07-04 2397
项目链接：https://mileyan.github.io/pseudo_lidar/ 代码链接：https://github.com/mileyan/pseudo_lidar 摘要 3D目标检测是一个重要的任务在自动驾驶领域。最近高精度检测率技术清单中，提供3D输入数据被提供从精确但是昂贵的LiDAR（激光雷达）技术。方法基于便宜的单目或者立体图像数据，直到现在，结果全部都...
BAD SLAM: Bundle Adjusted Direct RGB-D SLAM
CSS360的博客
06-27 1640
开源代码：http://www.eth3d.net/ 摘要 即时定位和重建（SLAM）系统的一个关键组成是加入3D图评估和相机轨迹的优化。捆绑调整（BA）对于这个是黄金标准。由于在稠密RGB-D SLAM中大量的变量，先前的工作集中在BA近似。相反，在这篇文章中，我们提出一个新颖快速直接BA方案，来应用到一个实时稠密的RGB-D SLAM 算法。 另外，我们展示...
Video Enhancement with Task-Oriented Flow
CSS360的博客
07-15 1553
项目地址：http://toflow.csail.mit.edu/ 摘要 一些视频增强算法依靠光流来存储帧在一个视频序列中。精确光流估计然而是很难处理的；以及光流它自己本身是一个局部最优的展现对于特别视频处理任务。在这篇文章中，我们提出了任务导向流（TOFlow），一个动作展示学习在一个自监督，特定任务方式。我们设计了一个神经网络使用一个训练的动向估计组成和一个视频处理组成，以...
ArcFace: Additive Angular Margin Loss for Deep Face Recognition
CSS360的博客
07-05 1494
open source：https://github.com/deepinsight/insightface 摘要 一个主要的挑战在特征学习，使用深度卷积神经网络（DCNNs）对大尺度人脸识别，是设计合适loss函数，它能够增强区别。中心损失距离在深度特征和他们相应的类中心在欧几里得空间去实现类内紧度。球体猜测这个线性转换矩阵在最后全卷积层，能够使用作为一个表示类中在欧几里得空...
Monocular Total Capture: Posing Face, Body, and Hands in the Wild ∗
CSS360的博客
07-19 1292
论文地址：http://openaccess.thecvf.com/content_CVPR_2019/papers/Xiang_Monocular_Total_Capture_Posing_Face_Body_and_Hands_in_the_CVPR_2019_paper.pdf 项目地址：https://github.com/CMU-Perceptual-Computing-Lab/Mon...
FDB Static Entries Static FDB entries are created by the CPU through the Address Update from CPU mechanism (Section 10.4.7, CPU Update and Query of the FDB). Static FDB entries are skipped by the FDB aging process (Section 10.4.10, FDB Entry Refresh/Aging), however static entries still have the <age/refresh> bit set when the Source Address lookup matches the entry. Static entries are optionally subject to the FDB delete and transplant mechanisms (Section 10.4.11, Deleting FDB Entries, Section 10.4.12, FDB Unicast Entry Transplanting). If a source MAC FDB lookup matches a static MAC entry, but the packet is received on a different location than that of the entry, the address does not move, and the packet is assigned to a globally configured command. The command options are:  FORWARD  MIRROR  TRAP  HARD DROP  SOFT DROP If the packet is assigned a DROP command, a global configuration determines if is considered a security breach event (Section 10.6.5, Bridge Security Breach Events), and is counted by the Bridge Drop Counter (Section 10.15.1.1, Bridge Drop Counter). Configuration  To configure if a source MAC Address received on a location other than the static FDB entry configuration, constitutes a security breach event, use the <Move Static Address Is Security Breach> field in the Bridge Global Configuration0 Register (Table 319 p. 2216).  To configure the packet command when a source MAC Address received on a location other than the static FDB entry configuration, use the <Move Static Command> field in the Bridge Command Configuration0 Register (Table 322 p. 2223 翻译一下并解释
最新发布
11-18
<think> Hmm, the user is asking for a translation and explanation of FDB static entries mechanism. This seems to be a continuation of previous technical discussions about FDB operations, indicating the user is likely a network engineer or hardware developer working with switch ASICs. The content describes how static FDB entries behave differently from dynamic ones - they bypass aging but still get refreshed, can be subject to delete/transplant mechanisms, and have special handling when packets arrive from wrong ports. Key technical nuances I need to clarify: First, the core distinction between static and dynamic entries lies in aging immunity, yet static entries still update their refresh bit during lookups. This refresh mechanism is interesting - it maintains activity tracking without actual aging. Second, the security aspects are critical. When static entries receive packets from unauthorized ports, the configurable actions (DROP/TRAP etc.) and optional security breach triggering create flexible policy enforcement. The dual-layer control through <Move Static Command> and <Move Static Address Is Security Breach> fields allows decoupling of packet handling from security logging. Third, the note about static entries participating in delete/transplant mechanisms suggests they aren't entirely immutable. This deserves emphasis since "static" typically implies permanence, but here they can still be moved or deleted under specific conditions. The user's technical level appears advanced given they're consistently asking about FDB internals. I should maintain precise register references (Table 322 etc.) while explaining behavioral implications. The previous questions about MAC learning and AU messages suggest they're building comprehensive understanding of FDB management. For the translation, I'll keep technical terms like "static FDB entries" untranslated but render all descriptions in natural Chinese. The explanation will focus on operational scenarios - perhaps contrasting static vs dynamic entry behavior through concrete examples would help. </think>以下是关于FDB静态表项的翻译及技术解析： --- ### **FDB静态表项机制** #### **核心特性** ```markdown 1. **创建方式** - 仅支持CPU通过AU消息创建（详见10.4.7节） 2. **老化豁免** - 跳过自动老化流程（10.4.10节） - 仍会更新`<age/refresh>`位（当源MAC匹配时） 3. **特殊操作** - 可配置是否参与删除/移植机制（10.4.11/10.4.12节） ``` #### **安全校验规则** 当数据包**源MAC匹配静态表项但入口位置不符**时： ```mermaid graph TB A[数据包到达] --> B{位置匹配静态表项？} B -->|是| C[正常转发] B -->|否| D[执行全局命令] D --> E[FORWARD/MIRROR/TRAP/DROP] E --> F{DROP命令？} F -->|是| G[触发安全事件计数] F -->|否| H[结束] ``` --- ### **关键配置寄存器** | **寄存器名称** | **字段** | **功能说明** | |-----------------------------------|--------------------------------------|---------------------------------------------| | `Bridge Command Configuration0` | `<Move Static Command>` | 配置位置不匹配时的包处理命令 | | `Bridge Global Configuration0` | `<Move Static Address Is Security Breach>` | 设置DROP是否触发安全违规事件（1=触发） | --- ### **设计思想解析** 1. **静态表项双重属性** - **持久性**：豁免老化 → 保障关键路径稳定性 - **活性跟踪**：刷新`<age/refresh>`位 → 监控业务流量状态 2. **安全隔离机制** - **位置绑定**：强制校验入口端口/Trunk - **弹性策略**：通过寄存器分离"动作执行"与"事件告警" *示例：可配置DROP但不触发安全事件，用于静默阻断* 3. **异常处理优势** ```markdown - 防止静态表项被非法篡改：错误位置直接阻断 - 避免配置错误导致环路：DROP命令立即丢弃问题流量 - 细粒度审计：安全事件计数精准定位攻击源 ```
评论 1
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
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
实付元
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