Research map / Label-efficient and robust learning / Recognition and long-tail
Get the Best of Both Worlds: Improving Accuracy and Transferability by Grassmann Class Representation
Haoqi Wang, Zhizhong Li, Wayne Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace vector class representation with 8-dim Grassmann class subspaces | ResNet50-D · ImageNet-1K val | top-1 accuracy: 78.04 -> 79.26 | Improved |
| replace linear fc with Grassmann fc across architectures | 8 architectures · ImageNet-1K val | top-1 accuracy: +0.44 to +1.38 | Improved |
| use multi-center representations (SoftTriple, SubCenterArcFace) instead of vector class | ResNet50-D · ImageNet-1K val | top-1 accuracy: 78.04 -> 75.55 (SoftTriple), 77.10 (SubCenterArcFace) | Hurt |
| increase subspace dimension k from 1 to 32 (ablation) | GCR ResNet50-D · linear transfer avg | accuracy (%): 73.64 -> 79.70 | Improved |
| increase subspace dimension beyond 16 (ablation) | GCR ResNet50-D · ImageNet-1K val | top-1 accuracy: 79.21 (k=16) -> 78.63 (k=32) | Hurt |
| normalize features to constant length (ablation) | GCR k=8 · ImageNet-1K val | top-1 accuracy: 78.12 -> 79.26 | Improved |
| use Riemannian SGD instead of vanilla SGD (ablation) | GCR k=8 · ImageNet-1K val | top-1 accuracy: 78.55 -> 79.26 | Improved |
| train backbone jointly instead of fine-tuning only the Grassmann layer (ablation) | GCR k=8 · ImageNet-1K val | top-1 accuracy: 78.14 -> 79.44 | Improved |
| use CosineSoftmax/LabelSmoothing/Dropout instead of softmax for feature transfer | ResNet50-D · linear transfer avg | accuracy (%): 77.98 -> 73.72 (CosineSoftmax) | Hurt |
| gamma 30 instead of 25 (ablation) | GCR k=8 · ImageNet-1K val | top-1 accuracy: 79.26 -> 78.47 | Hurt |
| geodesic retraction versus QR retraction (ablation) | GCR Riemannian SGD · ImageNet-1K val | top-1 accuracy: 79.26 vs 79.13 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ArcFace — ArcFace: Additive Angular Margin Loss for Deep Face Recognition (CVPR 2019) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Building Deep Networks on Grassmann Manifolds (AAAI 2018 · similarity 0.89)
- Approximate Grassmannian Intersections: Subspace-Valued Subspace Learning (ICCV 2017 · similarity 0.88)
- Random features for Grassmannian kernel approximation with bounded rank-one projections (NeurIPS 2026 · similarity 0.88)
- Learning Topology-Driven Multi-Subspace Fusion for Grassmannian Deep Networks (AAAI 2026 · similarity 0.86)
- Matrix Manifold Neural Networks++ (ICLR 2024 · similarity 0.85)
- Does SGD really happen in tiny subspaces? (ICLR 2025 · similarity 0.83)
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.
