I have some questions about “Minimizing Instance Uncertainty to align the distributions of the labeled and unlabeled instances”

[Elaineok]

[yuantn]

你好，我认为在你的推导和分析中有一些问题：

1. theta_g 和 theta_f1/2 是非线性的，所以步骤 4 不太合理。

2. 在训练过程中确实会存在一个可以使 loss=0 的特征向量。

3. 最小化分布偏差这个目标是通过两个步骤实现的（最大化 和 最小化 示例不确定性，如图 2（a）所示），而不仅仅是最小化不确定性。偏差的减小是通过这样的一个反复的过程实现的而不是一个单步操作。

---

Hi, I think that there are some problems in your derivation and analysis:

1. theta_g and theta_f1/2 are non-linear, so the step 4 is not reasonable.

2. There is indeed a feature vector, which can make loss=0.

3. Minimizing the distribution bias is achieved by TWO steps (maximizing AND minimizing uncertainty, as shown in Fig. 2(a) ) but not only minimizing uncertainty. The reduction of bias is achieved in such an iterative process but not a single step.