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

Tackling Dimensional Collapse toward Comprehensive Universal Domain Adaptation

ICML 2025 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Feature alignment for unsupervised adaptation · 0 citations (OpenAlex)

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.

Shows that partial domain matching (PDM) fails in extreme universal domain adaptation due to target dimensional collapse, and fixes it by adding a SimSiam-style self-supervised de-collapse loss on unlabeled target data, improving UAN, CMU, UniOT and MLNet.

Problem

Partial domain matching for universal domain adaptation underperforms the source-only baseline when many source-private classes are present (extreme UniDA), and the paper seeks to explain and fix this across the full spectrum of label-set distributions.

Approach

Identifies dimensional collapse of target representations as the cause of PDM failure under high source-private ratio, and proposes adding a self-supervised de-collapse loss (uniformity and alignment, instantiated with SimSiam) on unlabeled target data to preserve representation structure and improve importance-weight quality.

Key results

• In extreme UniDA on Office-Home (P2R), SO reaches 63.2 H-score while UAN drops to 27.3; UAN+SSL reaches 72.3 with SimSiam.
• UAN+SSL improves average H-score on Office31 from 48.1 (UAN) to 79.8 in the general setting... and on Office-Home (5/10/50) UniOT+SSL averages 77.1 vs 76.6 for UniOT.
• UniOT+SSL reaches 40.2 vs 38.3 on DomainNet S2R and MLNet+SSL averages 48.9 vs 47.8 on DomainNet in extreme setting.

Measured effects reported in the paper

Surprising findings

• All SOTA PDM methods (UAN, CMU, UniOT, DANCE, MLNet) fall below the simple source-only baseline in the high source-private-ratio regime.
• Uniformity loss alone resolves dimensional collapse, while alignment loss alone provides little benefit.

Limitations

• Improvement is smaller for self-training methods than for adversarial training methods.
• Evaluation limited to image classification benchmarks with ResNet-50 backbone.

Experimental setup

Datasets: Office31 Office-Home VisDA DomainNet

Benchmarks: Office31 Office-Home VisDA DomainNet

Metrics: H-score

Baselines: UAN CMU UniOT DANCE MLNet SO (source only)

Builds on: UAN CMU UniOT DANCE SimSiam

Closest papers by meaning

• Unified Optimal Transport Framework for Universal Domain Adaptation (NeurIPS 2022 · similarity 0.91)
• Universal Semi-Supervised Domain Adaptation by Mitigating Common-Class Bias (CVPR 2024 · similarity 0.90)
• Mutual Nearest Neighbor Contrast and Hybrid Prototype Self-Training for Universal Domain Adaptation (AAAI 2022 · similarity 0.90)
• Prototypical Partial Optimal Transport for Universal Domain Adaptation (AAAI 2023 · similarity 0.90)
• Gradient-Based Sample Selection for Black-Box Universal Domain Adaptation (AAAI 2025 · similarity 0.88)
• Category-Aware Active Domain Adaptation (ICML 2024 · similarity 0.88)