Methodology [论文原图,侵删]
文章将一组environment-user pair看成一个domain, 并没有利用Few shot learning
Fig 1. Train Model
Model input
A matrix with dimension , where
is the number of sampling points in the frequency domain, and
is the transceiver link number. So the input DFS series are with dimension
.
Feature Extractor
Fig 2. Architecture of feature extractor G
CNNs: extracts spatial features from each single DFS profile
GRUs: models the temporal dependencies of the whole sequence
为什么选择GRU:和LSTM相比paras更少!可有效防止过拟合。We choose GRUs because they are capable of learning longterm representations and they achieve comparable performance while involving fewer parameters compared to long short-term memory (LSTM).
希望提取的 能有一定的domain invariant 特性(可以通过后面的adversarial learning实现)
Domain Discriminator
文章针对每一个source domain 都设计了一个discriminator
(总共N个),每个都有同样的结构:two fully connected layers where the first layer has leaky Relu activation while the second layer has sigmoid activation to get the output. 来自target domain的
会通过全部 D 但是来自
-th source domain的只会通过对应的discriminator
。
domain discriminators 的outputs 反映了target domain 和 source domain 之间的similarities. If the output is close to 0.5, it means that the domain discriminator is uncertain about whether the sample is from source or target.
Gesture Recognizer
Fig 3. Components of gesture recognizer
同样针对每一个source domain 都设计了一个recognizer, 每个recognizer包含以下部分
1. Feature Disentanglement: 之前提取的特征 可能依旧包含一些domain related features (location/orientation etc), 需要进一步的将其分离成 class-related features(gesture classification
) 和 domain-related features (domain info regression
)
2. Gesture Classifier: 将获得的class-related features 进行后续分类
3. Domain Factor Regressor: 将获得的domain-related features 输入并计算出和 ground-truth
location and orientation 的 regression loss
4. Feature Reconstructor: 确保 extracted class-related features 和 domain-related features 可以 恢复成 original feature (MSE loss)