ReNC Improves Model Adaptation in Open-World Scenarios.

Traditional Test-Time Adaptation (TTA) methods, which aim to bridge the gap between source and target data domains, often fail when the label distribution changes significantly, a common challenge in "open-world" scenarios. This means models struggle to adapt when encountering entirely new classes or a different prevalence of known classes at test time. Researchers propose a new method called Reliable Neural Collapse approximation (ReNC) to tackle Open-World Test-Time Adaptation (OWTTA). ReNC utilizes the principle of neural collapse, a phenomenon where features of the same class converge to their class mean, and class means become maximally separable, as a structural prior. Guided by this prior, ReNC uses pre-trained classifier weights as initial prototypes for the source domain. It then filters out-of-distribution (OOD) samples by measuring their similarity to these prototypes, ensuring that only relevant samples contribute to adaptation. A neural collapse approximation mechanism refines these prototypes, allowing them to adapt to the target domain while preserving the beneficial neural collapse structure, leading to superior performance on open-world benchmarks.