- 分享主题:Multi-source Domain - 论文标题:Unsupervised Multi-source Domain Adaptation Without Access to Source Data - 论文链接:https://openaccess.thecvf.com/content/CVPR2021/papers/Ahmed_Unsupervised_Multi-Source_Domain_Adaptation_Without_Access_to_Source_Data_CVPR_2021_paper.pdf - 分享人:唐共勇 |
---|
1. Summary
【必写】,推荐使用 grammarly 检查语法问题,尽量参考论文 introduction 的写作方式。需要写出
- 这篇文章解决了什么问题?
- 作者使用了什么方法(不用太细节)来解决了这个问题?
- 你觉得你需要继续去研究哪些概念才会加深你对这篇文章的理解?
The standard unsupervised multi-source domain adaptation (UDA) uses the source data and the model trained on the source to perform adaptation in the target domain. On the contrary, the author of the paper introduced a setting that can adapt to multiple models without accessing the source data. The algorithm proposed by the author is based on pseudo labeling and information maximization. Using intuitive theory shows that the proposed framework is superior to the best available source and minimizes the impact of negative migration. To solve the problem of multi-source model adaptation without accessing source data, the author deploys the information maximization loss on the weighted combination of all source models. It solves the problem of multi-source unsupervised domain adaptation and no access to data. Propose a UDA algorithm that does not need to access source domain data. The algorithm generates the target model by optimizing a designed unsupervised loss and automatically identifying the optimal mixed source model. Under intuitive assumptions, the author establishes a theoretical guarantee for the performance of the target model, which shows that it is always at least as good as deploying a single optimal source model, to minimize negative migration
2. 你对于论文的思考
需要写出你自己对于论文的思考,例如优缺点,你的takeaways
Weighted Pseudo-labeling
由于领域漂移(domain shift),信息最大化可能会导致某些实例被错误的类集群所绑定由于领域的转移,信息最大化可能会导致某些实例被错误的类集群所绑定。这些错误的预测会在训练过程中得到强化,并导致一种被称为确认偏差(confirmation bias)的现象。作者为了抑制这种效果,采用了一种自监督聚类策略,该策略灵感来自于DeepCluster技术。
首先,计算每个源模型对整个目标数据集的聚类中心:
根据每个源模型的当前聚合权值,这些特定于源的中心将进行组合:
接下来,通过将每个样本分配到特征空间中最接近的聚类中心来计算其伪标签:
重复这个过程以获得更新后的中心和伪标签:
其中 1(·) 是一个指示函数,当自变量为真时给出值为1。虽然这种计算集群中心和伪标签的交替过程可以重复多次,以获得固定的伪标签,但一次就足够了。然后得到了这些伪标签的交叉熵损失
3. 其他
【可选】
参考blog
https://blog.csdn.net/singxsy/article/details/120319488