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

Memory-Assisted Sub-Prototype Mining for Universal Domain Adaptation

ICLR 2024 · 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.

Proposes MemSPM, a memory-assisted module that mines sub-prototypes within each category to handle intra-class concept shift in universal domain adaptation, combined with DCC on a frozen CLIP ViT-B/16 encoder.

Problem

Universal domain adaptation methods align each category with a single one-hot prototype, ignoring intra-class concept shift between samples of the same category, which hurts adaptation and unknown-class detection.

Approach

A learnable memory of N items, each partitioned into S sub-prototypes, retrieves top-K cosine-similar sub-prototypes via attention with an adaptive threshold to form a task-oriented embedding fed to the classifier; a decoder reconstructs images to visualize sub-prototypes, and DCC-style cycle-consistent alignment is applied on top.

Key results

• MemSPM+DCC reaches 84.2% average H-score on Office-Home UniDA, surpassing DCC (74.4%) and GLC (82.6%) with CLIP.
• On Office-31 OSDA, MemSPM+DCC obtains 95.6% H-score, a 5.5% gain over GLC.
• On DomainNet UniDA, MemSPM+DCC reaches 56.7% average H-score, versus 55.1% for GLC (ImageNet) and 52.2% for DCC (CLIP).

Measured effects reported in the paper

Surprising findings

• Using an ImageNet-pretrained ViT-B/16 instead of CLIP drops MemSPM to 76.7% H-score on Office-Home, and a randomly initialized encoder to 64.3%, showing strong dependence on pretraining.
• The ResNet-50 backbone makes memory training difficult, as early-stage embeddings lack discriminability and address only a few sub-prototypes.

Limitations

• Training the memory unit is hard with ResNet-50 and needs a strong pretrained encoder such as CLIP ViT-B/16.
• Number of sub-prototypes per memory item may need manual adjustment for category diversity.
• Results on some settings (e.g., PDA on Office-Home) are not uniformly best; PDA performance is only comparable to tailored methods.

Experimental setup

Datasets: Office-31 Office-Home VisDA DomainNet

Benchmarks: Office-31 Office-Home VisDA DomainNet

Metrics: H-score accuracy OS* UNK

Baselines: UAN CMU DCC OVANet UMAD GATE UniOT GLC ROS OSBP ETN BA3US

Builds on: DCC GLC CLIP OVANet Domain Consensus Clustering

Closest papers by meaning

• Concept-Based Unsupervised Domain Adaptation (ICML 2025 · similarity 0.87)
• Gradient-Based Sample Selection for Black-Box Universal Domain Adaptation (AAAI 2025 · similarity 0.87)
• Category-Aware Active Domain Adaptation (ICML 2024 · similarity 0.86)
• Target Semantics Clustering via Text Representations for Robust Universal Domain Adaptation (AAAI 2025 · similarity 0.86)
• Tackling Dimensional Collapse toward Comprehensive Universal Domain Adaptation (ICML 2025 · similarity 0.86)
• ADU: Adaptive Detection of Unknown Categories in Black-Box Domain Adaptation (CVPR 2025 · similarity 0.86)