Research map / Label-efficient and robust learning / Recognition and long-tail
Neuromanifold-Regularized KANs for Shape-fair Feature Representations
Mazlum Ferhat Arslan, Weihong Guo, Shuo Li
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 |
|---|---|---|---|
| add hybrid basis (different degrees in early and later layers) (ablation) | VGG-Gram CKAN · Tiny ImageNet cue-conflict | shape bias: +5.1% | Improved |
| add hybrid basis (ablation) | VGG-Gram CKAN · Tiny ImageNet val | top-1 accuracy: -3.8% | Hurt |
| add fused features from two branches (ablation) | VGG-Gram CKAN + hybrid basis · Tiny ImageNet cue-conflict | shape bias: +6.0% | Improved |
| add fused features from two branches (ablation) | VGG-Gram CKAN + hybrid basis · Tiny ImageNet val | top-1 accuracy: -4.4% | Hurt |
| add style decorrelation loss (ablation) | NMR-KAN · Tiny ImageNet cue-conflict | shape bias: +3.7% | Improved |
| add style decorrelation loss (ablation) | NMR-KAN · Tiny ImageNet val | top-1 accuracy: +0.9% | Improved |
| add hybrid basis (ablation) | VGG-Gram CKAN · Tiny ImageNet-C | relative mCE: 105.7 -> 100.1 | Improved |
| add fused features (ablation) | VGG-Gram + hybrid basis · Tiny ImageNet-C | relative mCE: 100.1 -> 95.1 | Improved |
| add style decorrelation loss (ablation) | NMR-KAN · Tiny ImageNet-C | relative mCE: 95.1 -> 92.5 | Improved |
| vary uniform activation degree across all layers (2 to 7) (ablation) | VGG-Gram CKAN · Tiny ImageNet cue-conflict | shape bias: 41.5 -> 42.0 | No clear effect |
| use heterogeneous degrees (7,3) in a single extractor (ablation) | VGG-Gram CKAN · Tiny ImageNet cue-conflict | shape bias: 42.9% vs baseline 42.5 | Improved |
| use heterogeneous degrees (2,3) or (3,2) (ablation) | VGG-Gram CKAN · Tiny ImageNet cue-conflict | shape bias: 47.65% shape bias | Improved |
Closest papers by meaning
- On the expressiveness and spectral bias of KANs (ICLR 2025 · similarity 0.85)
- Few-Shot Unsupervised Implicit Neural Shape Representation Learning with Spatial Adversaries (ICML 2024 · similarity 0.85)
- Estimating Shape Distances on Neural Representations with Limited Samples (ICLR 2024 · similarity 0.85)
- KAN: Kolmogorov–Arnold Networks (ICLR 2025 · similarity 0.84)
- $\texttt{NAISR}$: A 3D Neural Additive Model for Interpretable Shape Representation (ICLR 2024 · similarity 0.83)
- U-KAN Makes Strong Backbone for Medical Image Segmentation and Generation (AAAI 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.
