. No theoretical proof and explanation. A1. Just the lack of proof or theoretical explanation should not be a

We will address the concerns. Domain adaptation papers without proofs have been accepted by NeurIPS (e.g., We need to decide whether a sample is "known" or "unknown". Importantly, we do not even know whether we have "unknown" samples in target domain for the universal domain Then, even though there are many "unknown" samples or none of them, the objective function for "unknown" The entropy of a classifier output shows the confidence of the prediction. Such distance should be effective metric for "unknown" score under different proportions B, and C can be put closer. To perform well, features have to be well-clustered.