Research map / Architectures and efficiency / CNN design
Adaptive Aggregation of Monte Carlo Augmented Decomposed Filters for Efficient Group-Equivariant Convolutional Neural Network
Wenkai Zhao, Barbara Daria Wichtmann, Steffen Albert, Angelika Maurer, Frank Gerrit Zöllner, Jürgen Hesser
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 shear augmentation range of 0.12pi to FB basis (ablation) | WMCG ResNet18 5x5 · ImageNet40 | error (%): 19.60 -> 18.80 | Improved |
| add shear augmentation range of 0.25pi to FB basis (ablation) | WMCG ResNeXt50 5x5 · ImageNet40 | error (%): 27.60 -> 26.95 | Improved |
| increase shear range from 0.12pi to 0.25pi on MH basis (ablation) | WMCG ResNet18 3x3 MH · ImageNet40 | error (%): 24.55 -> 26.85 | Hurt |
| use high-order (8th) Fourier-Bessel basis instead of 1st (ablation) | WMCG ResNet18 single FB basis · ImageNet40 | error (%): 28.65 -> 55.45 | Hurt |
| use MC affine sampling instead of MC scale-only sampling (16 samples, width 1/16) (ablation) | MCG-CNN ResNet18 · ImageNet40 | error (%): 36.75 -> 32.65 | Improved |
| use MC instead of equally spaced sampling (scale, 16 samples, width 1/4) (ablation) | MCG-CNN ResNet18 · ImageNet40 | error (%): 19.00 -> 19.20 | No clear effect |
| replace conv layers with WMCG 5x5 layers | ResNet50 (Pixmix) · ImageNet1k-C | mCE (%): 54.23 -> 53.04 | Improved |
| replace 7x7 conv layers with WMCG | ConvNeXt-S · ImageNet40 | top-1 accuracy (%): 84.75 -> 86.65 | Improved |
| replace 7x7 conv layers with dynamic convolution | ConvNeXt-S · ImageNet1k | top-1 accuracy (%): 83.14 -> 82.57 | Hurt |
| replace conv layers with WMCG 5x5 layers | DudeNeXt-B · color denoising average | PSNR (dB): 31.35 -> 31.45 | Improved |
| enlarge WMCG kernel from 5x5 to 7x7 (ablation) | DnNeXt-B · grayscale denoising average | PSNR (dB): 29.26 -> 29.20 | Mixed |
| replace conv layers with WMCG | ResNeXt41 · STL10 OOD | accuracy (%): 81.13 -> 83.71 | Improved |
Closest papers by meaning
- Efficient Equivariant Network (NeurIPS 2021 · similarity 0.91)
- Group Downsampling with Equivariant Anti-aliasing (ICLR 2025 · similarity 0.90)
- Variational Partial Group Convolutions for Input-Aware Partial Equivariance of Rotations and Color-Shifts (ICML 2024 · similarity 0.88)
- Continuous Rotation Group Equivariant Network Inspired by Neural Population Coding (AAAI 2024 · similarity 0.88)
- Deep Ridgelet Transform and Unified Universality Theorem for Deep and Shallow Joint-Group-Equivariant Machines (ICML 2025 · similarity 0.87)
- Affine Equivariant Networks Based on Differential Invariants (CVPR 2024 · similarity 0.87)
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.
