AdaTriplet-RA: Domain matching via adaptive triplet and reinforced attention for unsupervised domain adaptation

Unsupervised domain adaptation (UDA) is a transfer learning task where the annotations of the source domain are available, but only have access to the unlabelled target data during training. Previous methods minimise the domain gap by performing distribution alignment between the source and target domains, which has a notable limitation, i.e., at the domain level, but neglecting the sample-level differences, thus preventing the model from achieving superior performance. To solve this, we improve the UDA task with an inter-domain sample-level matching scheme. We apply the widely-used Triplet loss to match the inter-domain samples. To reduce the catastrophic effect of the inaccurate pseudo-labels generated during training, we propose a novel uncertainty measurement method and use this uncertainty to select reliable pseudo-labels automatically. As the selection is uncertainty-aware, the pseudo labels are progressively refined as the training is performed. We apply the advanced Gumbel Softmax technique to realise an adaptive Top-k scheme to achieve adaptive selection. To enable the global ranking optimisation within one batch for the domain matching, the whole model is optimised via a reinforced attention mechanism, using the Average Precision (AP) of the domain matching as the reward. Our model AdaTriplet-RA achieves State-of-the-art results on several public benchmark datasets, and its effectiveness is validated via comprehensive ablation study. Our method improves the accuracy of the baseline by 9.7% (using ResNet-101 as the backbone network) and 6.2% (ResNet-50) on the VisDa dataset and 4.22% (ResNet-50) on the DomainNet dataset. The source code is publicly available at: https://github.com/shuxy0120/AdaTriplet-RA.