Open
Description
Hi authors,
Thanks for your inspiring work. I have a question and hope you can give me some insights.
In Section 4.2, the domain descriptor orthogonality (DDO) loss is designed to suppress the influence of domain-specific concepts simulated from possible unseen domains. Intuitively, this seems to focus on erasing the impact of concepts that do not yet exist in the training data. Could you clarify how suppressing these simulated unseen-domain-specific activations—rather than those actually learned from the training domain—helps improve generalization to unseen domains?
Best,
Nannan
Metadata
Metadata
Assignees
Labels
No labels