Research map / Label-efficient and robust learning / Domain adaptation and generalization

Batch Singular Value Polarization and Weighted Semantic Augmentation for Universal Domain Adaptation

ICML 2024 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Conditional distribution alignment

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes BSP-WSA for universal domain adaptation, combining an adversarial unknown-class classifier, a batch singular value polarization loss on predicted label matrices, and a class-weighted semantic augmentation loss, evaluated on Office-31, Office-Home and VisDA-2017.

Problem

Universal domain adaptation must detect target-private (unknown) classes while avoiding misclassifying target samples into source-private classes, under both domain and category shift.

Approach

Perform SVD on batch-level (memory-bank smoothed) predicted label matrices and maximize large singular values while minimizing small ones (singular value polarization) to suppress assignments to source-private classes; combine with a weighted version of semantic augmentation that scales per-class augmentation strength by estimated target-to-source sample ratios so that common classes are emphasize

Key results

• In the OPDA setting, BSP-WSA averages 77.0% H-score on Office-Home versus 75.6% for the second-best GATE, and 88.3% on Office-31 with 57.1% on VisDA S2R.
• In the PDA setting, BSP-WSA reaches 93.7%→95.0% on Office-31 and 74.6% average on Office-Home, outperforming GATE by about 2.4 points on average across the three datasets.
• Ablation on OPDA shows BSP alone raises Office-31 average from 70.5 to 86.3 and WSA alone to 80.1; both together give 88.3.

Measured effects reported in the paper

Surprising findings

• Forcing the BSP loss degrades performance under the ODA setting (no source-private classes), so only WSA helps there.

Limitations

• Relies on the assumption that target samples rarely fall into source-private classes and on a discriminative classifier and accurate unknown detection.
• Treats all target unknown classes as a single category.
• Gains are smaller in the ODA setting and on Office-31 ODA where no improvement over the best baseline is observed.

Experimental setup

Datasets: Office-31 Office-Home VisDA-2017

Benchmarks: Office-31 Office-Home VisDA-2017

Metrics: H-score

Baselines: UAN CMU DANCE DCC OVANet GATE GLC

Builds on: OVANet Semantic data augmentation (Li et al. 2021) Cui et al. SVD-based class diversity loss GLC

Closest papers by meaning

• Universal Semi-Supervised Domain Adaptation by Mitigating Common-Class Bias (CVPR 2024 · similarity 0.88)
• Gradient-Based Sample Selection for Black-Box Universal Domain Adaptation (AAAI 2025 · similarity 0.88)
• Tackling Dimensional Collapse toward Comprehensive Universal Domain Adaptation (ICML 2025 · similarity 0.87)
• Category-Aware Active Domain Adaptation (ICML 2024 · similarity 0.87)
• Low Category Uncertainty and High Training Potential Instance Learning for Unsupervised Domain Adaptation (AAAI 2024 · similarity 0.87)
• ADU: Adaptive Detection of Unknown Categories in Black-Box Domain Adaptation (CVPR 2025 · similarity 0.86)