Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Provably Uncertainty-Guided Universal Domain Adaptation

Yifan Wang, Lin Zhang, Ran Song, Paul L. Rosin, Yibin Li, Wei Zhang

TL;DR

The paper addresses UniDA by introducing an uncertainty-guided framework that avoids relying on potentially misaligned classifier outputs. It estimates the unknown probability $p(y=C+1\mid \mathbf{z})$ from target neighborhoods, discovers unknown samples via a linear-subspace NSLS approach with a $\delta$-filter, and trains with an uncertainty-guided margin loss plus an unknown-sample entropy loss and a supervised contrastive loss on the source. Key contributions include a Bayes-based empirical uncertainty estimate, a robust NSLS/$\delta$-filter unknown-discovery scheme, and an uncertainty-balanced learning objective, all validated by state-of-the-art results on Office-31, OfficeHome, and VisDA under OPDA and ODA settings. The method demonstrates strong generalization to different backbones (ResNet-50 and VGG) and provides insights for reducing domain misalignment while controlling overconfidence on unknowns, with practical impact for reliable UniDA in real-world tasks.

Abstract

Universal domain adaptation (UniDA) aims to transfer the knowledge from a labeled source domain to an unlabeled target domain without any assumptions of the label sets, which requires distinguishing the unknown samples from the known ones in the target domain. A main challenge of UniDA is that the nonidentical label sets cause the misalignment between the two domains. Moreover, the domain discrepancy and the supervised objectives in the source domain easily lead the whole model to be biased towards the common classes and produce overconfident predictions for unknown samples. To address the above challenging problems, we propose a new uncertainty-guided UniDA framework. Firstly, we introduce an empirical estimation of the probability of a target sample belonging to the unknown class which fully exploits the distribution of the target samples in the latent space. Then, based on the estimation, we propose a novel neighbors searching scheme in a linear subspace with a $δ$-filter to estimate the uncertainty score of a target sample and discover unknown samples. It fully utilizes the relationship between a target sample and its neighbors in the source domain to avoid the influence of domain misalignment. Secondly, this paper well balances the confidences of predictions for both known and unknown samples through an uncertainty-guided margin loss based on the confidences of discovered unknown samples, which can reduce the gap between the intra-class variances of known classes with respect to the unknown class. Finally, experiments on three public datasets demonstrate that our method significantly outperforms existing state-of-the-art methods.

Provably Uncertainty-Guided Universal Domain Adaptation

TL;DR

The paper addresses UniDA by introducing an uncertainty-guided framework that avoids relying on potentially misaligned classifier outputs. It estimates the unknown probability from target neighborhoods, discovers unknown samples via a linear-subspace NSLS approach with a -filter, and trains with an uncertainty-guided margin loss plus an unknown-sample entropy loss and a supervised contrastive loss on the source. Key contributions include a Bayes-based empirical uncertainty estimate, a robust NSLS/-filter unknown-discovery scheme, and an uncertainty-balanced learning objective, all validated by state-of-the-art results on Office-31, OfficeHome, and VisDA under OPDA and ODA settings. The method demonstrates strong generalization to different backbones (ResNet-50 and VGG) and provides insights for reducing domain misalignment while controlling overconfidence on unknowns, with practical impact for reliable UniDA in real-world tasks.

Abstract

Universal domain adaptation (UniDA) aims to transfer the knowledge from a labeled source domain to an unlabeled target domain without any assumptions of the label sets, which requires distinguishing the unknown samples from the known ones in the target domain. A main challenge of UniDA is that the nonidentical label sets cause the misalignment between the two domains. Moreover, the domain discrepancy and the supervised objectives in the source domain easily lead the whole model to be biased towards the common classes and produce overconfident predictions for unknown samples. To address the above challenging problems, we propose a new uncertainty-guided UniDA framework. Firstly, we introduce an empirical estimation of the probability of a target sample belonging to the unknown class which fully exploits the distribution of the target samples in the latent space. Then, based on the estimation, we propose a novel neighbors searching scheme in a linear subspace with a -filter to estimate the uncertainty score of a target sample and discover unknown samples. It fully utilizes the relationship between a target sample and its neighbors in the source domain to avoid the influence of domain misalignment. Secondly, this paper well balances the confidences of predictions for both known and unknown samples through an uncertainty-guided margin loss based on the confidences of discovered unknown samples, which can reduce the gap between the intra-class variances of known classes with respect to the unknown class. Finally, experiments on three public datasets demonstrate that our method significantly outperforms existing state-of-the-art methods.
Paper Structure (31 sections, 4 theorems, 35 equations, 11 figures, 4 tables)

This paper contains 31 sections, 4 theorems, 35 equations, 11 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Partial-set Domain Adaptation
  4. Open-set Domain Adaptation
  5. Universal Domain Adaptation
  6. Out-of-Distribution Detection
  7. Method
  8. Empirical Estimation of Unknown Samples
  9. Discovering Unknown Samples Based on the Uncertainty Estimation
  10. Neighbors searching in linear subspace (NSLS)
  11. $\delta$-filter for discovering unknown samples
  12. Learning
  13. Uncertainty-guided margin loss (UGM)
  14. Loss for unknown samples
  15. Supervised contrastive loss on source domain
  16. ...and 16 more sections

Key Result

Proposition 1

With the feature set of samples from two domains $Z^s$, $Z^t$ and a target feature $\mathbf{z} \in Z^t$, denoting $\hat{p}_{C\!+\!1}(\mathbf{z};k) = c_1\mathbbm{l}\{\max_{i=0,\dots,C}{\hat{p}_i(\mathbf{z};k,k_i\mid y=i)}\leq \beta\}$ where $k$ and $k_i$ are the number of neighbors and the number of Then, where $k_{max}=max_{i=0,\dots,C}(k_i)$, $\gamma\in [0,1]$, $r_k$ is the $k$-nearest neighbor

Figures (11)

  • Figure 1: (a) Illustration of Universal domain adaptation. (b) Comparison of UniDA methods. The distribution of embedings in the original feature space are highly misaligned because of the domain discrepancy. The existing neighborhood-based methods has bad performance on matching the samples from the two domains. Our method can reduce the influence of domain misalignment and find the unknown samples reliably.
  • Figure 2: The overall workflow of the proposed UniDA framework. By projecting the features extracted from the samples of both domains into a linear subspace, we estimate their uncertainty based on the neighbors searching and find the known/unknown samples. We refine the label of the known samples with the $\delta$-filter and send all the target samples into the classifiers with pseudo labels. Based on the prediction of the unknown samples, we can compute the uncertainty-guided margin loss which balances the intra-class variances of both domains and leads to a reliable classification.
  • Figure 3: Illustration for the effect of NSLS. Extracting the principal linear subspace can reduce the overlap of $\hat{p}$ of known target samples and unknown target samples.
  • Figure 4: Computation of $\delta$: the target sample should be compact enough with its neighbors compared to the compactness of the source class. The terms $\lambda_{i0}$, $\lambda_{in^{\prime}+1}$, $\hat{\lambda}_{i0}$, $\hat{\lambda}_{in^{\prime}+1}$ represent the singular values.
  • Figure 5: Graphs of distributions of $k$-nearest neighbor distances of target samples in an early epoch on Office-$31$ with the OPDA setting. The first row represents the distribution in the original feature space and the second row represents that in the subspace. The red lines represent the distribution of unknown target samples while the green lines represent that of known target samples.
  • ...and 6 more figures

Theorems & Definitions (7)

  • Proposition 1
  • Proof
  • Lemma 2
  • Proof
  • Lemma 3
  • Proof
  • Corollary 4