Adaptive Pseudo-Label Refinement by Negative Ensemble Learning for Source-Free Unsupervised Domain Adaptation

The majority of existing Unsupervised Domain Adaptation (UDA) methods\npresumes source and target domain data to be simultaneously available during\ntraining. Such an assumption may not hold in practice, as source data is often\ninaccessible (e.g., due to privacy reasons). On the contrary, a pre-trained\nsource model is always considered to be available, even though performing\npoorly on target due to the well-known domain shift problem. This translates\ninto a significant amount of misclassifications, which can be interpreted as\nstructured noise affecting the inferred target pseudo-labels. In this work, we\ncast UDA as a pseudo-label refinery problem in the challenging source-free\nscenario. We propose a unified method to tackle adaptive noise filtering and\npseudo-label refinement. A novel Negative Ensemble Learning technique is\ndevised to specifically address noise in pseudo-labels, by enhancing diversity\nin ensemble members with different stochastic (i) input augmentation and (ii)\nfeedback. In particular, the latter is achieved by leveraging the novel concept\nof Disjoint Residual Labels, which allow diverse information to be fed to the\ndifferent members. A single target model is eventually trained with the refined\npseudo-labels, which leads to a robust performance on the target domain.\nExtensive experiments show that the proposed method, named Adaptive\nPseudo-Label Refinement, achieves state-of-the-art performance on major UDA\nbenchmarks, such as Digit5, PACS, Visda-C, and DomainNet, without using source\ndata at all.\n