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

A Theory of Learning Unified Model via Knowledge Integration from Label Space Varying Domains

CVPR 2025 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Conditional distribution alignment · 2 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.

Proposes a joint-error learning theory and UM method for multi-source semi-supervised open-set domain adaptation with varying label spaces, plus an attention-based feature generation module for source-free settings.

Problem

Domain adaptation when multiple source domains have disjoint label spaces, the target has a few labeled samples, and unlabeled target data may contain unknown classes, including source-free cases.

Approach

Derives a target error bound via joint error over multiple sources, combines it with log-sum-exp weighting and multi-class PU learning for unknown detection, and generates labeled source anchors with an attention-based feature generation (AFG) module so source data is not needed.

Key results

• On Office-Home (1-shot/3-shot), UM improves HOS over the best baseline by 3.6%/3.8% with source data available.
• UM+AFG in the source-free setting improves HOS by 7.1%/7.0% on Office-Home and 4.9%/6.1% on DomainNet.
• With a ViT-B/16 target backbone, UM+AFG source-free HOS on Office-Home reaches 73.0 average, approaching or exceeding source-based results.

Measured effects reported in the paper

Surprising findings

• Removing progressive unknown rejection (PUR) drops DomainNet HOS from 64.8 to 57.4 on the ablation (w/o PUR), showing the unknown rejection is critical for source-free detection.
• Cluster-based baselines (OSBP, MOSDANET, HyMOS) are eventually overtaken by simple multi-class PU learning (MPU) as labeled target samples increase.

Limitations

• Source-free baselines required pseudo-source data built from confident target samples labeled by pre-trained models, so comparisons are adapted rather than native.
• Relies on approximated labeling function assumptions and a multiplier on the unknown predictive discrepancy to prevent all targets being labeled unknown.
• Evaluated only on two image benchmarks with artificially constructed disjoint source label spaces.

Experimental setup

Datasets: Office-Home DomainNet

Benchmarks: Office-Home DomainNet

Metrics: HOS OS* UNK

Baselines: OSBP PGL ANNA PUJE MOSDANET HyMOS MPU

Builds on: PUJE MPU SimCLR gradient reversal layer generalized VC theory

Closest papers by meaning

• Universal Semi-Supervised Domain Adaptation by Mitigating Common-Class Bias (CVPR 2024 · similarity 0.89)
• Open-Set Heterogeneous Domain Adaptation: Theoretical Analysis and Algorithm (AAAI 2025 · similarity 0.87)
• Open-Set Graph Domain Adaptation via Separate Domain Alignment (AAAI 2024 · similarity 0.87)
• A Unified Framework for Heterogeneous Semi-supervised Learning (CVPR 2025 · similarity 0.87)
• A General Representation-Based Approach to Multi-Source Domain Adaptation (ICML 2025 · similarity 0.87)
• Distinguish Then Exploit: Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment (CVPR 2025 · similarity 0.87)