Enhanced Adversarial Domain Generation via Optimized Batch Normalization and Classifier
Keywords: Adversarial, Domain Generation, Optimized Batch Normalization, Classifier
Abstract: In the endeavor of domain generalization, the objective is to develop a classification model, utilizing multiple source domains, that can subsequently be generalized to unseen target domains. The crux of domain generalization lies in discerning and learning discriminative features that are invariant across domains. Techniques utilizing adversarial domain generalization are paramount in achieving invariant representations. To mitigate the aforementioned impediment, we introduce a novel methodology, termed Auxiliary Classifier in Adversarial Domain Generalization (ACADG). ACADG endeavors to augment the diversity within the source domain through the integration of an auxiliary classifier. By amalgamating standard task-related losses—such as cross-entropy loss for classification and adversarial loss for domain discrimination—the overarching aim is to ensure the acquisition of condition-invariant features across all source domains, while concurrently enhancing the diversity of source domains. Our model demonstrates significant generalization capacity, surpassing contemporaneous state-of-the-art domain generalization methodologies. In the context of mathematical formalization, consider that the aforementioned adversarial loss, auxiliary classifier, and task-related losses are represented with pertinent LaTeX symbols and equations, reflecting the intricate interdependencies and mathematical nuances underpinning the proposed methodology.
Primary Area: optimization
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8148
Loading