- 分享主题：Transfer Learning, Unsupervised Domain Adaptation, Explicit Feature Distribution Alignment, CV
- 论文标题：Generalized Domain Conditioned Adaptation Network
- 论文链接：https://arxiv.org/pdf/2103.12339.pdf
1.Summary
This is a paper about classification tasks. Now there is a source domain and a target domain. The source domain data is labeled and the target domain data is not labeled. So this is a unsupervised domain adaptation problem. The GDCAN model proposed in this paper is an improvement of DCAN model. When the two domains are very different, even though DCAN weakly shares the models in the domain conditioned channel attention module, some models are still shared. Therefore, GDCAN improves the domain conditioned channel attention module of DCAN. When the difference between the two domains is large, the two domains do not share the neural network at all. When the two domains are relatively close, the original DCAN processing method is maintained. This part is called adaptive channel attention module. In order to deepen my understanding of this paper, I can read some papers on domain adaptation.2.你对于论文的思考
这是一篇关于DA的文章，提出的GDCAN模型是在DCAN模型上进行了改进。DCAN解决的问题是防止浅层网络共享而妨碍特定于域的特征的提取，但是DCAN的domain conditioned channel attention module中，虽然第一个fc是独立的，但是第二个fc不是，这仍然可能会妨碍特定于域的特征的提取，因此GDCAN对这一部分进行了改进，替换为了adaptive channel attention module，当两个域的差异过大时，第二个fc也是独立提取特征的，差异比较小时，就保持DCAN的网络设置。3. 其他
3.1 需要解决的问题
解决DCAN的domain conditioned channel attention module存在的问题，防止当两个域差异过大时，仍然存在特定于域的特征无法有效提取的问题。3.2 GDCAN
3.2.1 Domain Conditioned Channel Attention Modul
这是DCAN中的做法，如下面的式子所示，源域和目标域的数据不共享第一个fc。
3.2.2 Adaptive Channel Attention Module
这是GDCAN中的做法，是对Domain Conditioned Channel Attention Modul的改进。
先用下面的式子计算两个域之间的差异：
m_heat代表两个域之间的差异，并于阈值λ相比，选择网络结构：
3.3 实验
（1）数据集：Office-Home
（2）特征可视化（利用了t-SNE降到了二维）
图中蓝色和红色代表不同的域，GDCAN效果很明显。
若有收获，就点个赞吧
0 人点赞
