Source-linked AI summary
Domain Conditioned Adaptation Network
Shuang Li, Chi Harold Liu, Qiuxia Lin, Binhui Xie, Zhengming Ding, Gao Huang, Jian Tang
TL;DR
Deep domain adaptation commonly aligns task-specific features while sharing convolutional layers, although large domain differences may make that assumption unsuitable for learning domain-specific features. DCAN adds domain-conditioned channel attention and high-level feature correction, and the authors report stronger performance than other deep adaptation methods, especially on difficult cross-domain tasks.
Problem
Deep adaptation methods generally align top task-specific layers while sharing convolutional layers, which may fail to capture domain-informative low-level features when visual patterns differ substantially.
Source: Introduction · p. 1Method
DCAN uses domain-conditioned channel attention to learn domain-specific convolutional features and feature correction blocks to align high-level distributions across domains.
Source: Introduction · p. 2Results
DCAN achieves better performance than other deep domain adaptation methods, especially on very tough cross-domain tasks.
Source: Conclusion · p. 7Takeaways & Limitations
The authors report that its components can be inserted into any layer of an original network with an uncomplicated loss function.
Source: Conclusion · p. 7DCAN retains the assumption that source and target data should share most network parameters because the domains have some correlation.
Source: Domain Conditioned Convolutional… · p. 3
Abstract
from arXiv · showhide
Tremendous research efforts have been made to thrive deep domain adaptation (DA) by seeking domain-invariant features. Most existing deep DA models only focus on aligning feature representations of task-specific layers across domains while integrating a totally shared convolutional architecture for source and target. However, we argue that such strongly-shared convolutional layers might be harmful for domain-specific feature learning when source and target data distribution differs to a large extent. In this paper, we relax a shared-convnets assumption made by previous DA methods and propose a Domain Conditioned Adaptation Network (DCAN), which aims to excite distinct convolutional channels with a domain conditioned channel attention mechanism. As a result, the critical low-level domain-dependent knowledge could be explored appropriately. As far as we know, this is the first work to explore the domain-wise convolutional channel activation for deep DA networks. Moreover, to effectively align high-level feature distributions across two domains, we further deploy domain conditioned feature correction blocks after task-specific layers, which will explicitly correct the domain discrepancy. Extensive experiments on three cross-domain benchmarks demonstrate the proposed approach outperforms existing methods by a large margin, especially on very tough cross-domain learning tasks.
Introduction
Deep domain adaptation aligns task-specific features while usually sharing convolutional layers across domains, an assumption that may limit learning when their visual patterns differ substantially. DCAN uses domain-conditioned channel attention for low-level domain-specific features and feature correction for high-level discrepancy, with reported state-of-the-art results on three cross-domain benchmarks.
Introduction
- Earlier deep adaptation methods align task-specific features using discrepancy or adversarial losses while assuming convolutional layers are universal across domains.The paper notes that top-layer loss modifications may have decayed gradients when backpropagated through very deep networks.Source: Introduction · p. 1
- When domain visual patterns differ greatly, shared convolutional layers may fail to capture domain-informative low-level features.The passage contrasts clean-background Amazon images with daily-life images containing complex backgrounds.Source: Introduction · pp. 1–2
- Figure 1 compares last-layer attention maps for source-only, DCAN without and with channel attention, and a target-label-trained model on Ar→Rw.The authors report that DCAN with channel attention attends more like the target-label-trained model, while top-layer alignment alone can highlight irrelevant objects.Source: Figure 1 · p. 1, Introduction · p. 2
- DCAN combines domain-conditioned channel attention for low-level domain-specific knowledge with feature correction to reduce high-level feature-distribution discrepancy.The feature correction block uses the target distribution as a reference.Source: Introduction · p. 2
- DCAN reports several state-of-the-art results across three cross-domain benchmarks, especially the large-scale DomainNet benchmark.The authors say it addresses challenging adaptation from small to large scales.Source: Introduction · p. 2
Related Work
DCAN addresses the limits of shared convolutional backbones by learning domain-conditioned channel attention at low levels and correcting feature discrepancies after task-specific layers. Its formulation combines source classification, target entropy minimization, feature alignment, and source-based correction regularization.
Related Work
- Prior deep DA methods align task-specific features but generally share one convolutional backbone; attention-based DA approaches described here focus on spatial rather than channel attention.The paper positions DCAN's convolutional channel adaptation as a distinct way to model inter-domain and domain-specific features.Source: Related Work · pp. 2–3
Preliminary and Motivation
- DCAN targets large domain shifts by learning domain-specific low-level features alongside inter-domain features, then explicitly minimizing high-level mismatch with correction blocks.The motivation includes examples such as clean-background Amazon images versus daily-life images with complex backgrounds.Source: Preliminary and Motivation · p. 3
Domain Conditioned Convolutional Learning
- Separate source and target routes learn domain-wise channel dependencies, while shared components retain transfer; channel-wise attention then recalibrates each domain's convolutional features.The channel-attention module is described as weakly shared and as requiring only a small amount of additional parameters and computation.Source: Domain Conditioned Convolutional… · pp. 3–4
Task-specific Feature Alignment via Domain Conditioned Feature Correction
- Task-specific correction blocks add a target residual and use MMD to align corrected target features with source features, including after softmax to transfer category correlations.A source regularizer compares each class with a random source subset to guide correction and address over-correction.Source: Task-specific Feature Alignment via… · p. 4
Overall Formulation of DCAN
- The overall objective combines source cross-entropy, target entropy minimization, task-specific feature-alignment losses, and correction regularization.The target entropy term is motivated as a way to increase discrimination when target labels are unavailable.Source: Overall Formulation of DCAN · pp. 4–5
Experiment
Across three benchmarks, DCAN is evaluated against recent adaptation methods and its ablations; results favor DCAN overall. Attention analyses show larger domain differences in later stages and on harder tasks.
Experimental Setup
- Evaluation covers Office-31, Office-Home, and DomainNet, using both domains’ training sets and reporting target test-set results.Each transfer task is evaluated by averaging three random experiments.Source: Experimental Setup · p. 5
Comparison Results
- DCAN improves over baselines by 5.2%, 3.2%, and 4.3% with ResNet-50, ResNet-101, and ResNet-152, respectively.These are average improvements on DomainNet.Source: Comparison Results · p. 6, Table 3 · p. 6
- On Office-31, DCAN has the best average accuracy and exceeds MDD by 2.6% on D→A and DSBN by 0.5% on W→A.The reported exceptions are DAAA on D→W and SymNets on A→D, where DCAN is slightly lower.Source: Comparison Results · p. 6
- On Office-Home, DCAN achieves the highest average accuracy among compared methods, at 70.5%.The paper reports significant improvements over previous methods.Source: Comparison Results · p. 6
Empirical Analysis
- Removing both correction regularizers lowers average performance by 1.4%, while removing channel attention degrades adaptation performance by 2.0%.The study also reports that the full method outperforms the ablation variants.Source: Empirical Analysis · pp. 6–7, Table 4 · p. 6
- DCAN’s t-SNE features show better inter-class separation and intra-class clustering than ResNet-50 and DAN on A→W.ResNet-50 features are described as mismatched across domains, while DAN categories are not well aligned.Source: Empirical Analysis · p. 7, Figure 3 · p. 7
- Attention differences are smaller in the first three convolutional stages than in stage 4, indicating more domain-specific features in later stages.The authors infer that completely shared convolutional layers may be improper.Source: Empirical Analysis · p. 7, Figure 4 · p. 7
- At stage 4, Ar→Cl has a larger channel-attention difference than A→W, which the authors associate with the harder task’s greater domain discrepancy.The comparison is between Office-Home Ar→Cl and Office-31 A→W.Source: Empirical Analysis · p. 7, Figure 4 · p. 7
Conclusion
DCAN learns domain-specific convolutional features while mitigating mismatch in task-specific layers. It uses channel attention for low-level knowledge and feature-correction blocks to align high-level distributions.
Conclusion
- DCAN outperforms other deep domain-adaptation methods, especially on very tough cross-domain tasks.The conclusion reports better performance without giving specific benchmark values.Source: Conclusion · p. 7
- Domain-conditioned channel attention enriches low-level domain-specific knowledge to facilitate subsequent feature migration.The module supports simultaneous learning of domain-specific features in convolutional stages.Source: Conclusion · p. 7
- Feature-correction blocks align marginal and output distributions across domains, and each component can be inserted into network layers with an uncomplicated loss.Source: Conclusion · p. 7