Source-linked AI summary
Unsupervised Domain Adaptation via Structurally Regularized Deep Clustering
Hui Tang, Ke Chen, Kui Jia
TL;DR
UDA must predict unlabeled target data despite source–target distribution shift, while alignment-based transfer may damage target discrimination. The paper proposes SRDC, which clusters target data under source-based structural regularization, and reports that it outperforms existing methods without explicit domain alignment.
Problem
Alignment-based UDA can damage the intrinsic discrimination of target data, motivating methods that preserve or directly uncover that discrimination.
Method
SRDC uses deep discriminative clustering of target data with structural source regularization, enhanced by intermediate-feature clustering and soft selection of less divergent source examples.
Results
Without explicit domain alignment, SRDC outperforms all existing methods on the benchmark datasets, while ablations verify the efficacy of its components.
Takeaways & Limitations
The paper demonstrates a UDA strategy that directly uncovers target discrimination instead of explicitly aligning domains.
Takeaways & Limitations
The approach assumes structural similarity between source and target domains, and unlabeled target data alone does not guarantee sensible clustering solutions.
Abstract
from arXiv · showhide
Unsupervised domain adaptation (UDA) is to make predictions for unlabeled data on a target domain, given labeled data on a source domain whose distribution shifts from the target one. Mainstream UDA methods learn aligned features between the two domains, such that a classifier trained on the source features can be readily applied to the target ones. However, such a transferring strategy has a potential risk of damaging the intrinsic discrimination of target data. To alleviate this risk, we are motivated by the assumption of structural domain similarity, and propose to directly uncover the intrinsic target discrimination via discriminative clustering of target data. We constrain the clustering solutions using structural source regularization that hinges on our assumed structural domain similarity. Technically, we use a flexible framework of deep network based discriminative clustering that minimizes the KL divergence between predictive label distribution of the network and an introduced auxiliary one; replacing the auxiliary distribution with that formed by ground-truth labels of source data implements the structural source regularization via a simple strategy of joint network training. We term our proposed method as Structurally Regularized Deep Clustering (SRDC), where we also enhance target discrimination with clustering of intermediate network features, and enhance structural regularization with soft selection of less divergent source examples. Careful ablation studies show the efficacy of our proposed SRDC. Notably, with no explicit domain alignment, SRDC outperforms all existing methods on three UDA benchmarks.
1. Introduction
The paper proposes SRDC to uncover target-domain discrimination through discriminative clustering while using source labels for structural regularization, avoiding explicit domain alignment. It reports benchmark efficacy and superiority over existing methods without explicit alignment.
- Motivation: Mainstream UDA aligns source and target features so source-trained classifiers can predict target labels, but this may damage target discrimination.The paper frames this as a risk of alignment-based transfer.
- Motivation: Structural domain similarity combines domain-wise discriminative clustering with class-wise geometric closeness across domains.These assumptions motivate regularizing target clustering with labeled source data.
- Method: SRDC minimizes KL divergence between network predictions and an auxiliary distribution, replacing that distribution with source ground-truth labels through joint network training.It additionally clusters intermediate features and softly selects less divergent source examples.
- Method: SRDC directly uncovers target discrimination through discriminative clustering constrained by structural source regularization.The method is designed to avoid relying on explicit feature alignment.
- Results: Without explicit domain alignment, SRDC outperforms all existing methods on the benchmark datasets.Careful ablation studies also verify the efficacy of its individual components.
2. Related works
Related UDA methods primarily pursue domain-level or class-level alignment, along with approaches for discriminative feature learning, transferable-region selection, and data generation.
- Alignment based domain adaptation: Domain-adversarial methods align source and target domains as a whole so source class labels can transfer to unlabeled target data.This is described as domain-level alignment.
- Alignment based domain adaptation: Other methods minimize domain shift using metrics such as maximum mean discrepancy, while class-level methods align class centroids across domains.Centroid alignment uses labeled source and pseudolabeled target examples for shared classes.
- Discriminative feature learning: Some approaches detect non-discriminative features with domain-specific classifiers and learn discriminative feature extractors instead.Other methods focus on transferable regions or generate synthetic images with GANs.
3. The strategies of transferring versus uncovering the intrinsic target discrimination
The paper contrasts alignment-based transfer with directly uncovering target discrimination, grounding its approach in structural similarity between source and target domains. Alignment may damage target structure and reduce inductive UDA effectiveness.
- UDA setting: Transductive UDA predicts labels for unlabeled target examples, whereas inductive UDA evaluates learned models on held-out target instances.The distinction matters because inductive UDA requires models that remain consistent across source domains.
- Structural similarity: Structural similarity assumes intrinsic discriminative clusters within each domain and geometric closeness between clusters sharing a class label.These are termed domain-wise discrimination and class-wise closeness.
- Transferring versus uncovering: Alignment-based transfer learns shared representations for source-trained classifiers but risks damaging intrinsic target discrimination.The paper links this damage to reduced effectiveness in inductive UDA when adapted classifiers deviate from the oracle target classifier.
- Transferring versus uncovering: The proposed alternative directly uncovers intrinsic target discrimination through target discriminative clustering and constrains it with structural source regularization.Recent clustering methods still align domains through cluster centers or samples, retaining the identified risk.
4. Discriminative target clustering with structural source regularization
SRDC directly clusters unlabeled target instances while regularizing those clusters with labeled source supervision, rather than explicitly aligning domains. It further strengthens target discrimination through feature-space clustering and source regularization through similarity-based source weighting.
- Deep discriminative target clustering: SRDC minimizes KL divergence between the network's predictive label distribution and an auxiliary target distribution for discriminative clustering.The method alternates auxiliary-distribution updates with network updates that use the auxiliary distribution as labels.
- Deep discriminative target clustering: The target-clustering objective combines KL divergence with a balancing term that discourages merged clusters and encourages approximately uniform target cluster sizes.The balancing term is used because target label-distribution knowledge is unavailable.
- Deep discriminative target clustering: Feature-space clustering adds soft assignments based on distances between intermediate target features and learnable cluster centers.The same clustering objective is optimized while cluster centers are initialized and updated during training.
- Deep discriminative target clustering: Target clustering alone may not yield sensible solutions because its auxiliary distribution is unconstrained; source regularization serves a similar role to reconstruction-based regularization.Feature-space clustering is also described as alleviating this issue through soft assignments.
- Structural source regularization: Structural source regularization jointly trains the same network on labeled source examples by replacing the target auxiliary distribution with ground-truth source labels.Source supervision is implemented through cross-entropy minimization and shares the classifier's feature-space partition across domains.
- Enhancement via soft source sample selection: SRDC weights source examples by their similarity to target cluster centers, and experiments report improved results with these weighted objectives.The similarity is computed against the target center corresponding to each source example's class label.
5. Experiments
Experiments evaluate SRDC on three UDA benchmarks, analyze its components and source-sample selection, and compare it with existing methods. SRDC consistently improves transfer performance, including difficult tasks, without explicit domain alignment.
- 5.1. Setups: Experiments use Office-31, ImageCLEF-DA, and Office-Home under the standard UDA protocol with labeled source and unlabeled target samples.Office-31 has six transfer tasks, ImageCLEF-DA has six, and Office-Home has twelve.
- 5.2. Ablation studies and analysis: Removing structural source regularization, feature discrimination, or soft source sample selection degrades performance on Office-31.The ablation study evaluates each component using ResNet-50.
- 5.3. Comparisons with the state of the art: SRDC outperforms compared methods on almost all Office-31 transfer tasks, with especially strong gains on difficult A→W and W→A tasks.The comparison includes the latest BSP method targeting discriminability in adversarial feature adaptation.
- 5.3. Comparisons with the state of the art: SRDC achieves much better results than all compared methods on every ImageCLEF-DA transfer task, including difficult C→P and P→C tasks.These tasks involve source and target domains with equal size and class balance.
- 5.3. Comparisons with the state of the art: SRDC significantly exceeds compared methods on most Office-Home tasks and improves over MDD despite the benchmark’s greater visual and categorical difficulty.The paper notes that Office-Home still leaves substantial room for improvement.
6. Conclusion
SRDC uses structural domain similarity to directly uncover target discrimination through source-regularized deep discriminative clustering. It avoids explicit domain alignment while incorporating intermediate-feature clustering and soft selection of less divergent source examples.
- SRDC addresses potential damage to intrinsic target discrimination from alignment-based UDA by directly clustering target data with structural source regularization.
- The method minimizes KL divergence between network predictions and an auxiliary distribution, replacing that distribution with source ground-truth labels through joint network training.
- SRDC enhances target discrimination through clustering of intermediate network features and strengthens regularization by softly selecting less divergent source examples.
- Experiments on benchmark datasets testify to the efficacy of SRDC.
A. Other implementation details
SRDC jointly trains a feature embedding, classifier, and learnable cluster centers under its structurally regularized deep clustering objective. Training initializes target assignments with K-means and repeatedly updates network quantities using source and target mini-batches.
- SRDC simultaneously learns the feature embedding function θ, classifier ϑ, and learnable cluster centers {µ_k} under the structurally regularized deep clustering objective.
- Cluster centers are re-initialized at the start of each training epoch using current target cluster assignments.
- At the first epoch, target class assignments are obtained by standard K-means on embedded target features.
- K-means target cluster centers are initialized as source-data class centroids, while source-example weights are initialized to 1.
- The training algorithm samples target and source mini-batches and iterates network updates until convergence across training epochs.
B.1. Comparisons on Office-31
On Office-31 with ResNet-50, SRDC outperforms compared methods on almost all transfer tasks, supporting its effectiveness in this benchmark setting.
- SRDC outperforms almost all compared methods across Office-31 transfer tasks using ResNet-50.
B.2. Comparisons on ImageCLEF-DA
On ImageCLEF-DA with ResNet-50, SRDC exceeds every compared method, including CAN, on all transfer tasks by a large margin.
- SRDC exceeds all compared ImageCLEF-DA methods, including CAN, on every transfer task by a large margin.
B.3. Comparisons on Office-Home
On Office-Home with ResNet-50, SRDC achieves much better results than compared methods, including CAN, on almost all transfer tasks.
- B.3. Comparisons on Office-Home: SRDC achieves much better results than all compared methods, including CAN, on almost all Office-Home transfer tasks.CAN results were obtained by running its official code.
- B.3. Comparisons on Office-Home: Table 8 reports Office-Home results as percentages using ResNet-50 as the base network.
- B.3. Comparisons on Office-Home: The comparison incorporates results quoted from the respective papers or prior works for existing methods.