Abstract
Unsupervised domain adaptation (UDA) aims to train a target classifier with labeled samples from the source domain and unlabeled samples from the target domain. Classical UDA learning bounds show that target risk is upper bounded by three terms: source risk, distribution discrepancy, and combined risk. Based on the assumption that the combined risk is a small fixed value, methods based on this bound train a target classifier by only minimizing estimators of the source risk and the distribution discrepancy. However, the combined risk may increase when minimizing both estimators, which makes the target risk uncontrollable. Hence the target classifier cannot achieve ideal performance if we fail to control the combined risk. To control the combined risk, the key challenge takes root in the unavailability of the labeled samples in the target domain. To address this key challenge, we propose a method named E-MixNet. E-MixNet employs enhanced mixup, a generic vicinal distribution, on the labeled source samples and pseudo-labeled target samples to calculate a proxy of the combined risk. Experiments show that the proxy can effectively curb the increase of the combined risk when minimizing the source risk and distribution discrepancy. Furthermore, we show that if the proxy of the combined risk is added into loss functions of four representative UDA methods, their performance is also improved.
Introduction
In Unsupervised Domain Adaptation (UDA), the objective is to leverage labeled data from a source domain to build a model that generalizes well to an unlabeled target domain. Theoretically, the error on the target domain is bounded by three critical terms: Source Risk (performance on source data), Distribution Discrepancy (the difference between source and target feature distributions), and Combined Risk (the minimum joint error achievable by a single classifier on both domains).
Historically, UDA research has focused almost exclusively on minimizing the first two terms, operating under the assumption that the combined risk is a negligible constant. However, this paper demonstrates that as models are forced to align disparate distributions, the combined risk is dynamic and can significantly increase. This "hidden" risk can cause target accuracy to degrade even when the distributions appear well-aligned.
The E-MixNet Framework
To address the unobservable nature of combined risk due to the lack of target labels, the authors introduce E-MixNet. This framework rethinks the theoretical bound, proving that when using neural feature transformers, the combined risk becomes a function of the transformation itself rather than a constant.
Enhanced Mixup (e-mixup)
The core of E-MixNet is the e-mixup strategy, which creates a proxy for the combined risk. Unlike standard mixup, e-mixup specifically pairs unlabeled target samples with "pure" labeled source samples. This "source anchoring" provides a stable ground-truth reference, preventing the model from drifting due to noisy target pseudo-labels. Furthermore, the strategy prioritizes pairing samples that are distant in the feature space to better explore the decision boundaries.
Proxy Loss and Optimization
A dedicated proxy classifier ($C^*$) is trained on these e-mixup samples using Mean Squared Error (MSE). MSE was chosen over standard cross-entropy because it is bounded and less sensitive to the "confidently wrong" predictions often encountered in UDA. The final training objective is a multi-objective minimax game:
- Minimize Source Risk.
- Minimize Distribution Discrepancy (often via a domain discriminator or GRL).
- Minimize the Combined Risk Proxy.
Experimental Evaluation
The researchers evaluated E-MixNet on three major UDA benchmarks: Office-31, Office-Home, and Image-CLEF.
Key Results and Insights
The empirical results confirm the theoretical intuition. As shown in the visualizations, standard methods often see a spike in combined risk as training progresses, which correlates with a drop in target accuracy. E-MixNet effectively curbs this increase, maintaining a low combined risk and achieving superior accuracy.
- State-of-the-Art Performance: E-MixNet achieved an average accuracy of 90.2% on Office-31 and 70.6% on Office-Home, outperforming strong baselines like MDD and SymNets.
- Universal Applicability: By integrating the combined risk proxy into existing methods like DAN, DANN, and CDAN, the authors observed consistent performance gains. For example, DANN's accuracy improved by 5.5% simply by accounting for this third term.
Conclusion
This work highlights that the "combined risk" is a vital but often ignored component of the UDA learning bound. By introducing E-MixNet and the e-mixup strategy, the authors provide a practical way to estimate and control this risk. While the method introduces additional hyperparameters and computational complexity, the significant and universal performance improvements across benchmarks suggest that controlling combined risk is essential for robust domain adaptation. Future work may extend this principle to more complex scenarios like partial or multi-source domain adaptation.
Summary updated