DOME: Learning Transferable Domain Variables from Sparse Supervision for Test-Time Adaptation
Researchers have developed DOME, a new method for test-time adaptation that explicitly models domain variables from sparse supervision. Unlike previous approaches that infer a single global domain distribution, DOME uses vision-language pretraining to extract dense representations and parameterize domains as distributional variables. This approach, which includes a momentum-updated sparse domain bank, enables even basic adaptation strategies to achieve state-of-the-art performance on various benchmarks, suggesting that explicit domain representation is key to robust adaptation. AI
IMPACT Explicit domain representation could lead to more robust AI models in real-world, shifting environments.