Source-linked AI summary
Dynamic Weighted Learning for Unsupervised Domain Adaptation
Ni Xiao, Lei Zhang
TL;DR
UDA seeks to classify unlabeled target data using labeled source data, but existing methods may inadequately balance domain alignment and class discriminability. DWL dynamically weights these objectives and reweights samples for domain-size imbalance, achieving strong results across benchmark tasks.
Problem
Existing UDA methods may ignore the interaction between domain alignment and class discrimination, making their tradeoff inadequate and leaving negative transfer possible.
Method
DWL monitors alignment and discriminability during training, dynamically weights their learning losses, and weights samples according to cross-domain sample imbalance.
Results
DWL reports 77.1% average accuracy on VisDA-2017, 97.3% on MNIST→USPS, 97.4% on USPS→MNIST, and strong results on difficult transfer tasks.
Takeaways & Limitations
Dynamic balancing and sample weighting support application across cross-domain classification scenarios with different feature types, sample sizes, and initial distributions.
Takeaways & Limitations
The theoretical setup assumes source and target domains have different marginal and conditional distributions.
Abstract
from arXiv · showhide
Unsupervised domain adaptation (UDA) aims to improve the classification performance on an unlabeled target domain by leveraging information from a fully labeled source domain. Recent approaches explore domain-invariant and class-discriminant representations to tackle this task. These methods, however, ignore the interaction between domain alignment learning and class discrimination learning. As a result, the missing or inadequate tradeoff between domain alignment and class discrimination are prone to the problem of negative transfer. In this paper, we propose Dynamic Weighted Learning (DWL) to avoid the discriminability vanishing problem caused by excessive alignment learning and domain misalignment problem caused by excessive discriminant learning. Technically, DWL dynamically weights the learning losses of alignment and discriminability by introducing the degree of alignment and discriminability. Besides, the problem of sample imbalance across domains is first considered in our work, and we solve the problem by weighing the samples to guarantee information balance across domains. Extensive experiments demonstrate that DWL has an excellent performance in several benchmark datasets.
1. Introduction
Unsupervised domain adaptation reduces reliance on costly labeled target data by transferring knowledge from labeled source domains. DWL dynamically balances alignment and class discriminability while weighting samples to address domain-size imbalance.
- Unsupervised domain adaptation transfers models from labeled source domains to unlabeled or partially labeled target domains by alleviating domain divergence.
- Deep adaptation methods commonly learn domain-invariant representations, while other approaches explicitly promote feature discriminability using LDA, clustering, or eigenvalue constraints.
- DWL monitors alignment and discriminability in real time and dynamically adjusts their learning weights during training.
- DWL weights samples according to the imbalance in source and target sample sizes to reduce training bias.
- The dynamic weighting mechanism is designed for cross-domain data with varied feature types and sample sizes.
2. Related Work
Related work addresses domain alignment and feature discriminability through diverse adaptation methods. However, the supplied passages highlight imbalance between these objectives and illustrate it using DANN and MCD.
- DDC uses a pretrained deep network with an adaptive layer based on Maximum Mean Discrepancy, while DANN uses adversarial learning for domain alignment.
- TPN aligns class-specific score distributions, BSP restricts eigenvalues, and ETD combines attention-aware transport distance with entropy-based regularization.
- Figure 1 measures alignment with MMD distance and discriminability with LDA’s max J(W), where smaller MMD and larger max J(W) indicate better outcomes.
3. Imbalance Between Alignment and Discrimination
The section shows that alignment and discriminability can move in opposite directions during adaptation. DANN and MCD therefore motivate dynamically controlling the tradeoff between the two objectives.
- DANN improves domain alignment while its class discriminability worsens during training on USPS→MNIST.
- For MCD, discriminability also declines after epoch 25 while MMD distance continues decreasing, indicating excessive alignment can produce negative transfer.
- Because both alignment and class discrimination affect final classification accuracy, the method should dynamically prevent either objective from worsening the other.
4. Theoretical Insight
The theoretical discussion frames target error through source error, domain divergence, and combined error, then motivates dynamically reducing divergence and combined error without increasing either term.
- UDA assumes labeled source samples and unlabeled target samples drawn from different marginal and conditional distributions.
- The target-error upper bound contains source error ϵS(h), H∆H-divergence dH∆H(Ds, Dt), and combined error λ of the ideal joint hypothesis.
- Source error is expected to be small because source labels are reliable, while alignment methods reduce domain discrepancy and adversarial methods generally treat λ as small.
- Excessive alignment may reduce dH∆H(Ds, Dt) while reducing discriminability and increasing λ, so λ is closely related to class discriminability.
- DWL seeks to decrease dH∆H(Ds, Dt) and λ synchronously by dynamically balancing alignment and discrimination learning.
5. Approach
DWL addresses sample imbalance and the interaction between domain alignment and class discriminability by weighting samples and dynamically balancing their learning losses. It monitors alignment and discriminability during training to reduce excessive alignment, excessive discriminability pursuit, and associated model bias.
- Sample Weighting: DWL weights samples from the two domains according to their sample-size imbalance to reduce training bias toward the larger domain.The sample-weighting degree is controlled by hyper-parameter a ∈ (0,1].
- Domain Alignment Learning and Class Discrimination Learning: The method uses a feature generator, domain discriminator, and three classifiers to learn aligned and discriminative representations.The classifiers are C1, C2, and C; the domain-alignment pathway uses adversarial learning, while class discrimination uses classifier discrepancy.
- Dynamic Weighted Learning: The balance factor τ is dynamically computed from normalized alignment and discriminability measures and updated during iterative training.τ approaches 0 when alignment is much better, 1 when alignment is much worse, and approximately 0.5 when the two degrees are equal.
- Dynamic Weighted Learning: MMD measures cross-domain alignment, while max J(W) based on LDA measures class discriminability, with larger max J(W) indicating better discriminability.The two evaluation values are min-max normalized to [0,1] before constructing the balance factor.
- Overall Training Objective: The overall minimax objective combines source cross-entropy, domain alignment loss weighted by τ, and class discrimination loss weighted by 1 − τ.This objective integrates sample weighting, alignment learning, discriminability learning, and dynamic weighting.
6. Experiments
Experiments across digits, object-recognition, and domain-adaptation benchmarks evaluate DWL against existing methods. DWL reports strong accuracy, improved feature alignment and clustering, and benefits from dynamic balancing and sample weighting.
- Benchmark results: DWL achieves 77.1% average accuracy on VisDA-2017, exceeding BSP's 75.9%.The method dynamically weights transferability and discriminability during iterative training.
- Benchmark results: DWL reaches 97.3% on MNIST→USPS and 97.4% on USPS→MNIST, outperforming state-of-the-art methods.The reported results are attributed to balancing domain alignment and class discrimination.
- Benchmark results: On Office-31 tasks D→A and W→A, DWL obtains 73.1% and 69.8%, versus ETD's 71.0% and 67.8%.These tasks involve a small source domain or large domain shift, respectively.
- Benchmark results: DWL outperforms CAT on ImageCLEF-DA, indicating that its learned feature space maintains class-conditional structure.The method is also described as plug-and-play for existing UDA methods.
- Feature visualization: t-SNE visualizations show source and target samples forming 10 clusters with clear boundaries as training progresses.The clustering effect is reported as good by epoch 150 on MNIST→USPS.
- Convergence analysis: Classification error and the balancing factor τ gradually converge to flat values during USPS→MNIST and MNIST→USPS training.The analysis links decreasing τ with greater emphasis on class discriminability and decreasing classification error.
- Ablation study: Ablation experiments show significant accuracy gains from sample weighting and dynamic τ, while dynamic weighting outperforms static loss-weight ratios.Sample weighting is tested on W→A, where domains contain 795 and 2817 samples.
7. Conclusions
DWL dynamically balances domain alignment and class discrimination for cross-domain classification while correcting bias from unequal domain sample sizes. Experiments show this dynamic learning idea is useful and scalable in domain adaptation.
- DWL dynamically adjusts domain-alignment and class-discrimination weights for different cross-domain classification scenarios.The method analyzes their interaction and balances them during learning.
- A sample-weighting mechanism addresses model bias caused by imbalanced sample sizes between domains.
- Extensive experiments demonstrate that dynamic learning between H∆H-divergence and combined error λ is useful and scalable in domain adaptation.