Source-Free Adaptation to Measurement Shift via Bottom-Up Feature Restoration

Source-free domain adaptation (SFDA) aims to adapt a model trained on\nlabelled data in a source domain to unlabelled data in a target domain without\naccess to the source-domain data during adaptation. Existing methods for SFDA\nleverage entropy-minimization techniques which: (i) apply only to\nclassification; (ii) destroy model calibration; and (iii) rely on the source\nmodel achieving a good level of feature-space class-separation in the target\ndomain. We address these issues for a particularly pervasive type of domain\nshift called measurement shift which can be resolved by restoring the source\nfeatures rather than extracting new ones. In particular, we propose Feature\nRestoration (FR) wherein we: (i) store a lightweight and flexible approximation\nof the feature distribution under the source data; and (ii) adapt the\nfeature-extractor such that the approximate feature distribution under the\ntarget data realigns with that saved on the source. We additionally propose a\nbottom-up training scheme which boosts performance, which we call Bottom-Up\nFeature Restoration (BUFR). On real and synthetic data, we demonstrate that\nBUFR outperforms existing SFDA methods in terms of accuracy, calibration, and\ndata efficiency, while being less reliant on the performance of the source\nmodel in the target domain.\n