当前位置:网站首页>ACM Multimedia 2022 | 视觉语言预训练模型中社会偏见的反事实衡量和消除
ACM Multimedia 2022 | 视觉语言预训练模型中社会偏见的反事实衡量和消除
2022-07-04 21:33:00 【智源社区】
论文标题:
Counterfactually Measuring and Eliminating Social Bias in Vision-Language Pre-training Models
论文链接:
https://arxiv.org/submit/4382207
代码和数据集:
https://github.com/VL-Bias/VL-Bias
在本工作中,我们针对VLP模型的多模态特性,基于反事实样本生成衡量了各个模态的社会偏见以及多模态融合的社会偏见。衡量预训练模型中的社会偏见本质上是衡量偏见概念(比如,性别等社会属性)和目标概念之间的相关性。
我们提出了基于反事实的偏见衡量方法CounterBias,对每个样本都生成偏见属性上的反事实样本,之后使用反事实前后模型对目标属性的建模变化来衡量偏见概念和目标概念的联系。
由于目前还没有现成的数据集专门用于分析VLP模型中的社会偏见,我们提出了VL-Bias数据集,用于本工作以及社区对VLP模型中的社会偏见的理解。VL-Bias包含24K图文对,其中目标属性包含52个动作和13个职业。在VL-Bias数据集上,使用CounterBias,我们测试了两种典型的VLP结构,单流VLP和双流VLP。关键的观测结果包括:(1)测试的VLP模型中存在社会偏见,视觉模态和文本模态中都存在;(2)包含在VLP模型中的社会偏见与人类刻板影响基本一致;(3)测试的单流架构VLP和双流架构VLP在两个模态中的偏见表现出不同的一致性;(4)测试的VLP模型表现出比BERT更强的性别偏见。
继承了CounterBias中的反事实思路,我们提出了一个简单但有效的方法FiarVLP消除VLP模型中的社会偏见。具体来说,我们首先在视觉和语言模态中分别生成偏见概念的反事实样本。再将MLM任务对事实和反事实的预测的概率差异最小化,从而防止模型学习偏差和目标概念之间的关联。实验结果证明了FairVLP的有效性。
边栏推荐
- MongoDB中的索引操作总结
- Go language loop statement (3 in Lesson 10)
- Flink1.13 SQL basic syntax (I) DDL, DML
- GTEST from ignorance to proficiency (3) what are test suite and test case
- Flutter WebView示例
- 做BI开发,为什么一定要熟悉行业和企业业务?
- gtest从一无所知到熟练使用(2)什么是测试夹具/装置(test fixture)
- AcWing 2022 每日一题
- [early knowledge of activities] list of recent activities of livevideostack
- What is business intelligence (BI), just look at this article is enough
猜你喜欢
![[leetcode] 17. Letter combination of telephone number](/img/be/7f456c092f7cda5ebabc2f1cce292e.png)
[leetcode] 17. Letter combination of telephone number

At the right time, the Guangzhou station of the city chain science and Technology Strategy Summit was successfully held

Arcgis 10.2.2 | arcgis license server无法启动的解决办法

Operation of adding material schedule in SolidWorks drawing

Word文档中标题前面的黑点如何去掉

解析互联网时代的创客教育技术

Why do you have to be familiar with industry and enterprise business when doing Bi development?

Master the use of auto analyze in data warehouse

How is the entered query SQL statement executed?

bizchart+slider实现分组柱状图
随机推荐
HDU - 1078 FatMouse and Cheese(记忆化搜索DP)
【C語言】符號的深度理解
[optimtool.unconstrained] unconstrained optimization toolbox
Go language loop statement (3 in Lesson 10)
At the right time, the Guangzhou station of the city chain science and Technology Strategy Summit was successfully held
Operation of adding material schedule in SolidWorks drawing
Lambdaquerywrapper usage
Relational database
更强的 JsonPath 兼容性及性能测试之2022版(Snack3,Fastjson2,jayway.jsonpath)
flink1.13 sql基础语法(一)DDL、DML
股票开户流程是什么?使用同花顺手机炒股软件安全吗?
开源之夏专访|Apache IoTDB社区 新晋Committer谢其骏
Basic structure of PostgreSQL - table
From repvgg to mobileone, including mobileone code
Master the use of auto analyze in data warehouse
面试官:说说XSS攻击是什么?
How was MP3 born?
Liu Jincheng won the 2022 China e-commerce industry innovation Figure Award
Cadeus has never stopped innovating. Decentralized edge rendering technology makes the metauniverse no longer far away
Redis has three methods for checking big keys, which are necessary for optimization