Uncertainty-guided Source-free Domain Adaptation
Lists
Only you see this
Abstract
Source-free domain adaptation (SFDA) aims to adapt a classifier<br> to an unlabelled target data set by only using a pre-trained source<br> model. However, the absence of the source data and the domain shift<br> makes the predictions on the target data unreliable. We propose quantifying<br> the uncertainty in the source model predictions and utilizing it to<br> guide the target adaptation. For this, we construct a probabilistic source<br> model by incorporating priors on the network parameters inducing a distribution<br> over the model predictions. Uncertainties are estimated by employing<br> a Laplace approximation and incorporated to identify target data<br> points that do not lie in the source manifold and to down-weight them<br> when maximizing the mutual information on the target data. Unlike recent<br> works, our probabilistic treatment is computationally lightweight,<br> decouples source training and target adaptation, and requires no specialized<br> source training or changes of the model architecture. We show the<br> advantages of uncertainty-guided SFDA over traditional SFDA in the<br> closed-set and open-set settings and provide empirical evidence that our<br> approach is more robust to strong domain shifts even without tuning.
BibTeX
@inproceedings{roy2023uncertainty,
author = {Subhankar Roy and Martin Trapp and Andrea Pilzer and Juho Kannala and Nicu Sebe and Elisa Ricci and Arno Solin},
title = {{Uncertainty-guided Source-free Domain Adaptation}},
booktitle = {Zenodo},
year = {2023},
doi = {10.5281/zenodo.7566109},
}Loading related content...