(有待完善)
创新点 参考 Few-Shot Adversarial Domain Adaptation
2.UDA的两个缺陷
首先,数据量小
二,不能确保不同domain但是同类别的data的语义对齐
3.1 Handling Scarce Target Data
问题一: target 的data少,怎么办,扩充它
by pairing them with each training source sample.
问题二:需要用到target的label信息
对策: 创造pairs of samples
domain-class discriminator DCD 判断是四种pair中的哪一个(3)
然后,更新**
g
t
g_t
gt** , 使得DCD不能区分group 1, 2,3,4(12 相同label 34 不同label)
the DCD can no longer distinguish between groups 1 and 2, and also between groups 3 and 4 using
the loss
Connection with multi-class discriminators:
(5)是本文的最终loss,既能实现class separability又能实现domain confusion。训练的时候还是min-max的思想
Training
训练DCD的时候freezing
g
s
g_s
gs,
g
t
g_t
gt
训练
g
s
g_s
gs,
g
t
g_t
gt ,
h
h
h 的时候freezing DCD
weight sharing
g s g_s gs, g t g_t gt 之间进行weight sharing