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

Unsupervised Continual Domain Shift Learning with Multi-Prototype Modeling

CVPR 2025 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Feature alignment for unsupervised adaptation · 1 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 Multi-Prototype Modeling (MPM), which keeps one domain-specific prototype classifier per domain plus a parameter-free bi-level graph enhancer, to improve unsupervised continual domain shift learning over domain-invariant methods.

Problem

Unsupervised continual domain shift learning (UCDSL): adapting a pre-trained model to a sequence of unlabeled target domains while generalizing to unseen domains and avoiding catastrophic forgetting, where existing domain-invariant methods suffer from an adaptivity gap.

Approach

Replace a single universal domain-invariant classifier with one prototype-based classifier per domain (Multi-Prototype Learning), theoretically tightening the error bound, and add a parameter-free Bi-Level Graph Enhancer that fuses domain-level prototype predictions via label propagation and calibrates predictions with category-level graph label propagation.

Key results

• On PACS, MPM reaches an All score of 84.9 versus 82.9 for CoDAG, averaged over 10 domain orders.
• On Digits-five, MPM achieves All 87.5 versus 85.7 for CoDAG.
• On DomainNet, MPM achieves All 68.6 versus 66.0 for CoDAG; ablations show MPL alone gives 78.2 average and full model 80.3 across datasets.

Measured effects reported in the paper

Limitations

• Only evaluated on classification; not adapted to regression tasks such as object detection.
• Theoretical analysis simplified to binary classification and two domains.
• Replay buffer limited to 200 items.

Experimental setup

Datasets: PACS Digits-5 DomainNet

Benchmarks: PACS Digits-five DomainNet

Metrics: TDA TDG FA All

Baselines: SHOT+ SHOT++ Tent AdaCon EATA L2D PDEN RaTP CoDAG

Builds on: CoDAG RaTP SHOT iCaRL

Closest papers by meaning

• Continual Learning of Domain-Invariant Representations (ICML 2026 · similarity 0.86)
• Subsidiary Prototype Alignment for Universal Domain Adaptation (NeurIPS 2022 · similarity 0.86)
• Unsupervised Domain Adaptation With Similarity Learning (CVPR 2018 · similarity 0.86)
• Domain Adaptation with Invariant Representation Learning: What Transformations to Learn? (NeurIPS 2021 · similarity 0.86)
• Tackling Dimensional Collapse toward Comprehensive Universal Domain Adaptation (ICML 2025 · similarity 0.85)
• A General Representation-Based Approach to Multi-Source Domain Adaptation (ICML 2025 · similarity 0.84)