Research map / Point clouds / Point cloud classification
On the Spectral Differences Between NTK and CNTK and Their Implications for Point Cloud Recognition
Yuanqu Mou, Chang Gou, Haiyang Bai, Jia Liu
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 |
|---|---|---|---|
| use 1D convolutional NTK instead of MLP NTK for point clouds | NTK, L=7 · ModelNet10 (xyz only) | accuracy: 17.58 -> 76.65 | Improved |
| use 1D convolutional NTK instead of MLP NTK for point clouds | NTK, L=7 · ModelNet40 | accuracy: 11.10 -> 60.62 | Improved |
| use 2D convolutional NTK instead of MLP NTK for images | NTK, L=7 · MNIST | accuracy: 94.00 -> 96.00 | Improved |
| use 2D convolutional NTK instead of MLP NTK for images | NTK, L=7 · CIFAR10 | accuracy: 59.19 -> 76.79 | Improved |
| add MLP layers after convolutions (PointNTK vs 1dCNTK) (ablation) | 1dCNTK · ModelNet40 | accuracy: 60.62 -> 80.47 | Improved |
| add MLP layers after convolutions (PointNTK vs 1dCNTK) (ablation) | 1dCNTK · ModelNet10 (xyz+orientation, full train) | accuracy: 91.96 -> 91.19 | No clear effect |
| use PointNTK kernel regression instead of PointNet with 100 training samples | PointNet · ModelNet40 (100 train samples) | accuracy: 38.32 -> 45.71 | Improved |
| use PointNTK instead of PointNet on full training data | PointNet · ModelNet40 | accuracy: 88.71 -> 80.47 | Hurt |
| increase PointNTK depth from 7 to 15 layers (L) (ablation) | PointNTK · ModelNet40 | accuracy: 80.47 -> 82.50 | Improved |
| replace max pooling and CE loss by average pooling and least squares (PointNetm) (ablation) | PointNet · ModelNet40 (100 train samples) | accuracy: 38.32 -> 5.17 | Hurt |
| add global average pooling to 1dCNTK (ablation) | 1dCNTK · ModelNet10 (xyz only), D=5 | accuracy: 44.38 -> 76.87 | Improved |
| add global average pooling to supervised 1dCNN (ablation) | 1dCNN · ModelNet10 (xyz only), D=5 | accuracy: 42.18 -> 39.54 | No clear effect |
Closest papers by meaning
- Scaling Neural Tangent Kernels via Sketching and Random Features (NeurIPS 2021 · similarity 0.89)
- Learning Coordinate-based Convolutional Kernels for Continuous SE(3) Equivariant and Efficient Point Cloud Analysis (CVPR 2026 · similarity 0.85)
- Neural Tangent Kernels Motivate Cross-Covariance Graphs in Neural Networks (ICML 2024 · similarity 0.85)
- Parameter-Efficient Fine-Tuning in Spectral Domain for Point Cloud Learning (TPAMI 2025 · similarity 0.85)
- Extrapolation and Spectral Bias of Neural Nets with Hadamard Product: a Polynomial Net Study (NeurIPS 2022 · similarity 0.85)
- The Surprising Effectiveness of Infinite-Width NTKs for Characterizing and Improving Model Training (AAAI 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.
