Research map / Label-efficient and robust learning / Test-time adaptation
Reconstruct and Match: Out-of-Distribution Robustness via Topological Homogeneity
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 REMA, which combines slot-attention-based selective reconstruction with hypergraph-based matching to capture topological homogeneity for OOD generalization and test-time adaptation.
Problem
Out-of-distribution robustness: deep models degrade on OOD data, and existing invariance methods mostly learn pairwise alignments while ignoring high-order structure of objects.
Approach
Decompose images into a sparse set of discrete slot vectors via a selective slot-based reconstruction module, then model high-order part relations with hypergraph construction, hypergraph convolution, and Sinkhorn-based cross-domain hypergraph matching.
Key results
- REMA reaches 80.0% average accuracy on PACS, Office-Home, and VLCS, above the best listed baseline GVRT at 78.1%.
- On CIFAR-10C, CIFAR-100C, and ImageNet-C under test-time adaptation, REMA achieves an average error of 33.3%, versus 34.6% for TSD.
- Ablation on PACS shows SSR alone reaches 88.1%, HORR alone 89.0%, and both together 90.3%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use REMA instead of VNE | OOD generalization · PACS | accuracy (%): 86.9 -> 88.7±0.3 | Helped |
| use REMA instead of GVRT | OOD generalization · Office-Home | accuracy (%): 70.1 -> 72.0±0.4 | Helped |
| use REMA instead of GVRT | OOD generalization · VLCS | accuracy (%): 79.0 -> 79.4±0.3 | Helped |
| use REMA instead of TSD | test-time adaptation · CIFAR-10C | error rate (%): 13.1 -> 12.0 | Helped |
| use REMA instead of TSD | test-time adaptation · CIFAR-100C | error rate (%): 37.7 -> 35.8 | Helped |
| use REMA instead of TSD | test-time adaptation · ImageNet-C | error rate (%): 53.2 -> 52.2 | Helped |
| add SSR to REMA (ablation) | REMA · VLCS | accuracy (%): 78.3 -> 79.4 | Helped |
| add SSR to REMA (ablation) | REMA · PACS | accuracy (%): 89.0 -> 90.3 | Helped |
| add SSR to REMA (ablation) | REMA · Office-Home | accuracy (%): 70.2 -> 71.6 | Helped |
| add HORR to REMA (ablation) | REMA · VLCS | accuracy (%): 78.6 -> 79.4 | Helped |
| add HORR to REMA (ablation) | REMA · PACS | accuracy (%): 88.1 -> 90.3 | Helped |
| add HORR to REMA (ablation) | REMA · Office-Home | accuracy (%): 70.3 -> 71.6 | Helped |
Surprising findings
- Many baseline OOD generalization methods perform worse than ERM on Office-Home, indicating limited scalability.
Limitations
- Learned slots may not be readily interpretable as concepts due to lack of fine-grained annotations.
- Authors avoided stronger feature extractors such as DINOSAUR for fairness, so absolute performance may be understated.
Experimental setup
Datasets: PACS VLCS Office-Home CIFAR-10 CIFAR-100 ImageNet
Benchmarks: PACS VLCS Office-Home CIFAR-10C CIFAR-100C ImageNet-C
Metrics: accuracy average error rate
Baselines: ERM CORAL DANN MMD MixStyle SagNet VNE GVRT Tent SHOT T3A TAST
Builds on: Slot Attention hypergraph convolution Sinkhorn layer ERM Tent
Closest papers by meaning
- AnchorNet: A Weakly Supervised Network to Learn Geometry-Sensitive Features for Semantic Matching (CVPR 2017 · similarity 0.85)
- Towards Scalable Topological Regularizers (ICLR 2025 · similarity 0.85)
- Decomposing Out-of-Distribution Error in Conditional Flow Matching via Wasserstein Geometry (ICML 2026 · similarity 0.85)
- Self-Supervised Learning of Geometrically Stable Features Through Probabilistic Introspection (CVPR 2018 · similarity 0.84)
- Multi-Part Object Representations via Graph Structures and Co-Part Discovery (NeurIPS 2026 · similarity 0.84)
- OCRT: Boosting Foundation Models in the Open World with Object-Concept-Relation Triad (CVPR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.