Abstract
Universal domain adaptation (UniDA) has been proposed to transfer knowledge learned from a label-rich source domain to a label-scarce target domain without any constraints on the label sets. In practice, however, it is difficult to obtain a large amount of perfectly clean labeled data in a source domain with limited resources. Existing UniDA methods rely on source samples with correct annotations, which greatly limits their application in the real world. Hence, we consider a new realistic setting called Noisy UniDA, in which classifiers are trained with noisy labeled data from the source domain and unlabeled data with an unknown class distribution from the target domain. This paper introduces a two-head convolutional neural network framework to solve all problems simultaneously. Our network consists of one common feature generator and two classifiers with different decision boundaries. By optimizing the divergence between the two classifiers' outputs, we can detect noisy source samples, find "unknown" classes in the target domain, and align the distribution of the source and target domains. In an extensive evaluation of different domain adaptation settings, the proposed method outperformed existing methods by a large margin in most settings.
Introduction
Unsupervised Domain Adaptation (UDA) typically assumes that the source and target domains share an identical label space. However, real-world scenarios often involve mismatched class sets. Universal Domain Adaptation (UniDA) addresses this by allowing for "Partial DA" (source classes missing in target) and "Open-set DA" (target classes missing in source, or "unknowns").
A significant remaining hurdle is the assumption of clean source labels. In practice, large-scale source data is often crawled or crowdsourced, leading to significant label noise. This paper introduces Noisy Universal Domain Adaptation (Noisy UniDA), a framework designed to handle noisy source labels alongside mismatched class distributions.
As shown in Figure 1, the model must distinguish between three types of samples: common classes present in both domains, source-private classes, and target-private classes (unknowns). Crucially, it must also identify noisy source samples—common or private class samples that have been incorrectly labeled.
Methodology: Divergence Optimization
The proposed solution utilizes a two-head convolutional neural network (CNN) architecture. It features a shared feature generator ($G$) and two distinct classifiers ($F_1$ and $F_2$) initialized with different weights. The core mechanism is the optimization of "divergence"—the disagreement between these two classifiers.
The Disagreement Principle
The fundamental insight is that two independent classifiers trained on the same data will naturally agree on clean, common samples but disagree on noisy or "unknown" samples. This disagreement acts as a multi-purpose filter for both source noise and target unknowns.
Training Procedure
The training follows a four-step iterative process to refine the model's ability to separate these data types:
- Step A-1 (Source Initialization): The model is initialized using a "small-loss" selection strategy. It assumes clean samples are learned faster, so it only trains on source samples with low cross-entropy loss to mitigate the impact of noise.
- Step A-2 (Target Update): The whole network is updated using target samples where the joint divergence ($L_{JD}$) is greater than zero, helping to explore the target distribution.
- Step B (Classifier Maximization): The classifiers are trained to maximize divergence on "unknown" target samples. This forces the classifiers to produce different boundaries for samples they haven't seen in the source domain.
- Step C (Generator Minimization): The generator is trained to minimize divergence for common samples. This aligns the target distribution with the source distribution specifically for the shared classes, avoiding "negative transfer" from private classes.
Experimental Results
The method was rigorously tested on standard benchmarks (Office, OfficeHome, and VisDA) with synthetic label noise (Symmetry and Pair flipping) at 20% and 45% levels.
Robustness to High Noise
The proposed method demonstrated exceptional stability in high-noise environments where traditional UniDA methods failed. For instance, on the OfficeHome dataset with 45% symmetry noise, the model achieved 71.97% accuracy, significantly outperforming the next best method, DANCE, which reached 66.96%.
On the large-scale VisDA dataset, the gains were even more pronounced, often exceeding baselines by 10-15%. Ablation studies confirmed that the two-head divergence mechanism was the primary driver of this performance, as removing it led to a sharp decline in accuracy across all noise levels and class set configurations.
Conclusion
This work presents a robust framework for Noisy UniDA by leveraging divergence between two classifiers. By treating disagreement as a signal for both label noise and class mismatch, the model achieves state-of-the-art performance in realistic, "messy" data transfer scenarios. While the method relies on specific hyperparameters like selection ratios and entropy thresholds, its ability to maintain high accuracy under nearly 50% label noise marks a significant step toward practical, real-world domain adaptation.
Summary updated