当前位置:网站首页>NAACL 2022 | TAMT:通过下游任务无关掩码训练搜索可迁移的BERT子网络
NAACL 2022 | TAMT:通过下游任务无关掩码训练搜索可迁移的BERT子网络
2022-06-27 13:30:00 【智源社区】
论文标题:Learning to Win Lottery Tickets in BERT Transfer via Task-agnostic Mask Training

图2 TAMT在预训练任务上(MLM或知识蒸馏)学习子网络结构,然后将其迁移到不同的下游任务进行微调
基于以上动机,我们提出下游任务无关的掩码训练(Task-Agnostic Mask Training,TAMT)方法。如图 2 所示,TAMT 在预训练任务上优化 BERT 子网络的结构(不改变预训练参数值),从而使子网络在预训练任务上有较好的性能。随后搜索到的子网络将被迁移到多种下游任务进行微调训练。
边栏推荐
- 对半查找(折半查找)
- CMOS级电路分析
- Openfeign service interface call
- ensp云朵配置
- Journal quotidien des questions (6)
- 诗歌一首看看
- JVM performance tuning and monitoring tools -- JPS, jstack, jmap, jhat, jstat, hprof
- After the deployment is created, the pod problem handling cannot be created
- POSIX AIO -- Introduction to glibc version asynchronous IO
- 《预训练周刊》第51期:重构预训练、零样本自动微调、一键调用OPT
猜你喜欢
随机推荐
Differences in perspectives of thinking
防火墙基础之华为华三防火墙web页面登录
Pytorch learning 1 (learning documents on the official website)
MySQL locking mechanism and four isolation levels
芯片供给过剩之际,进口最多的中国继续减少进口,美国芯片慌了
Deploy redis sentinel mode using bitnamiredis Sentinel
基于SSM实现招聘网站
全球芯片市场或陷入停滞,中国芯片逆势扩张加速提升自给率
IJCAI 2022 | 用一行代码大幅提升零样本学习方法效果,南京理工&牛津提出即插即用分类器模块
SFINAE
Privacy computing fat offline prediction
基于 xml 配置文件的入门级 SSM 框架整合
[tcapulusdb knowledge base] Introduction to tcapulusdb tcapsvrmgr tool (III)
Prometheus 2.26.0 new features
IJCAI 2022 | greatly improve the effect of zero sample learning method with one line of code. Nanjing Institute of Technology & Oxford proposed the plug and play classifier module
[problem solving] which nodes are run in tensorflow?
POSIX AIO -- glibc 版本异步 IO 简介
Realization of hospital medical record management system based on JSP
诗歌一首看看
以前国产手机高傲定价扬言消费者爱买不买,现在猛降两千求售

![[WUSTCTF2020]girlfriend](/img/a8/33fe5feb7bcbb73ba26a94d226cc4d.png)







