Source free domain adaptation improves with mutual model correction
Safeguarding Mutual Correction in Source-Free Domain Adaptation via Cut Statistics
Artificial Intelligence
Summary
When a model trained on one type of data needs to work on a new kind of data, it can make mistakes. Usually, it tries to fix itself but often gets stuck in repeating errors. The authors found that using two different models that tend to make different mistakes can help each other correct errors. They developed a method that carefully decides when one model should teach the other, even without knowing the true answers. This approach helps both models improve, making the adaptation to new data more reliable.
What this means in practice
- •For machine learning engineers: Improve model adaptation to new unlabeled data by safely exchanging corrections between models without needing source data access.
- •For computer vision teams: Enhance accuracy in adapting pretrained vision models to different target environments using dual-model supervision guided by statistical reliability.
Authors
Seongjun Lee, Changhee Lee
Abstract
Source-Free Domain Adaptation (SFDA) aims to adapt a source-pretrained model to an unlabeled target domain without access to the original source domain. While early single-model approaches rely on self-refinement, they are inherently susceptible to confirmation bias and struggle to correct their own systematic errors. To overcome this limitation, recent methods introduce Vision-Language (ViL) models as external knowledge sources. However, these approaches operate in a largely unidirectional paradigm, using the ViL model primarily to supervise the source-pretrained model. This overlooks a key structural property: the two models exhibit distinct failure modes -- where one produces an incorrect prediction, the other may produce a correct one, creating a natural opportunity for mutual correction within the target domain. Yet, without ground-truth labels, identifying which model is correct on any given sample is non-trivial, and naively exchanging predictions risks propagating errors across models. To address this challenge, we propose SafeCut, a novel approach that leverages the cut statistic as a label-free measure of prediction reliability to gate cross-model supervision. Our approach dynamically controls both the direction and strength of supervision based on relative reliability, selectively amplifying true corrections while suppressing miscorrections on a per-sample basis. We further provide theoretical justification showing that this reliability-gated mechanism guarantees a net-positive correction signal. Extensive experiments across diverse SFDA benchmarks demonstrate that SafeCut achieves state-of-the-art performance, highlighting the effectiveness of safeguarding mutual correction in SFDA via cut statistics.