域对抗网络Domain adversarial neural network及其应用相关论文
- 论文介绍
- 1.Domain-Adversarial Training of Neural Networks, JMLR,2016
- 2. Developing Personalized Models of Blood Pressure Estimation from Wearable Sensors Data Using Minimally-trained Domain Adversarial Neural Networks, MLHC,2020
- 3. A Principled Approach for Learning Task Similarity in Multitask Learning, IJCAI, 2019.
- 4. Multi-domain adversarial learning, ICLR, 2019
- 5.Task Similarity Estimation Through Adversarial Multitask Neural Network, IEEE TNNLS, 2021.
- 6. A Novel Neural Network Model based on Cerebral Hemispheric Asymmetry for EEG Emotion Recognition | IJCAI18
- 7. A Bi-hemisphere Domain Adversarial Neural Network Model for EEG Emotion Recognition, IEEE Transactions on Affective Computing, 2018.
- 8. Multi-Domain Adversarial Learning for Slot Filling in Spoken Language Understanding
- 总结
论文介绍
1.Domain-Adversarial Training of Neural Networks, JMLR,2016
这篇论文是域对抗网络的开山之作,作者从domain adaption理论的角度出发, 最终提出了域对抗网络, 论文的数学性比较强,写的非常详细。回过头来看,其思路也非常简单: 在训练中为了学到与目标域更加general的特征,在中间特征层额外加了一个域分类器用来区分不同样本所属的域,一方面,域分类器通过最小化域分类损失来提高其判别能力,另一方面,通过最大化域分类损失来更新特征提取子网络的参数,这实际上就是一个min-max优化问题。
网络示意图如下: 主要由三个module组成: feature extractor,label predictor, domain classifier, 如图所示,域分类器的损失在 feature extractor层会有gradient reverserial 操作 (因为是最大化)。
参数更新规则
:
可以看到:feature extractor module的参数更新有点意思: 主任务的损失在该处反向传播以最小化主任务(可以是分类或者回归)损失,另一方面,域分类器的损失在此处正向传播以欺骗域分类器。从而使得学习到的特征:informative 并且 general/cross domain.
2. Developing Personalized Models of Blood Pressure Estimation from Wearable Sensors Data Using Minimally-trained Domain Adversarial Neural Networks, MLHC,2020
将域对抗的思想应用于血压预测(回归任务)
3. A Principled Approach for Learning Task Similarity in Multitask Learning, IJCAI, 2019.
多任务学习框架, 将域对抗的思想用于多任务学习中,使得学习到的特征更加general (cross-task), 作者还推导了多任务框架下利用任务间相似性的泛化误差界。值得一看。
4. Multi-domain adversarial learning, ICLR, 2019
作者推导了半监督场景下的多域对抗学习的泛化误差界,将其应用于生物医学图像数据集,好像是显微镜数据。论文值得一看。
5.Task Similarity Estimation Through Adversarial Multitask Neural Network, IEEE TNNLS, 2021.
将域对抗的思想用于多任务学习中, 估计任务相似度。
6. A Novel Neural Network Model based on Cerebral Hemispheric Asymmetry for EEG Emotion Recognition | IJCAI18
7. A Bi-hemisphere Domain Adversarial Neural Network Model for EEG Emotion Recognition, IEEE Transactions on Affective Computing, 2018.
8. Multi-Domain Adversarial Learning for Slot Filling in Spoken Language Understanding
总结
未完待续…