New OVDU protocol rigorously removes unwanted domains from vision-language models

Researchers have introduced Open-Vocabulary Domain Unlearning (OVDU), a new protocol designed to rigorously remove specific stylistic domains from vision-language models (VLMs). Unlike previous methods that only overfit to seen class-domain pairs, OVDU ensures that domain erasure is class-agnostic and transfers to unseen classes. The proposed solution involves a parameter-editing framework that uses Fisher Information to isolate domain-sensitive weights and Targeted Manifold Scattering to disrupt the domain's stylistic geometry. Experiments on PACS, OfficeHome, and DomainNet datasets demonstrate that OVDU significantly improves open-vocabulary generalization and achieves superior sample efficiency. AI

IMPACT This research could lead to more robust and controllable vision-language models, particularly in sensitive applications like medical AI and autonomous driving.

RANK_REASON The cluster describes a new research paper introducing a novel method for unlearning specific domains from VLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Approximate Domain Unlearning
• arXiv
• DomainNet
• Fisher information
• Hugging Face
• OfficeHome
• Open Vocabulary Domain Unlearning
• picture archiving and communication system
• Targeted Manifold Scattering
• vision-language model

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →