Research map / Architectures and efficiency / CNN design
SONIC: Spectral Oriented Neural Invariant Convolutions
Gijs Joppe Moens, Regina Beets-Tan, Eduardo H P Pooch
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 SonicNet instead of ConvNet under combined perturbation | SynthShape · combined perturbation 20 | Dice: 0.62 -> 0.76 | Improved |
| Use SonicNet instead of ViT under 1.50 rescaling | SynthShape · rescaling 1.50 | Dice: 0.68 -> 0.74 | Improved |
| Use SonicNet instead of NIFF under Gaussian noise 0.2 | SynthShape · Gaussian noise sigma 0.2 | Dice: 0.92 -> 0.79 | Hurt |
| Use SonicNet instead of ConvNet under 45 degree rotation | SynthShape · rotation 45 deg | Dice: 0.28 -> 0.24 | Hurt |
| Use SonicNet instead of ConvNet on HalliGalli | HalliGalli · HalliGalli | accuracy: 0.42 -> 1.00 | Improved |
| Add inference noise sigma=0.1 to HalliGalli (SonicNet) | SonicNet · HalliGalli | accuracy: 1.00 -> 0.86 | Hurt |
| Use SonicNet instead of NIFF under HalliGalli noise | HalliGalli · HalliGalli (noise sigma=0.1) | accuracy: 0.56 -> 0.86 | Improved |
| Use SonicNet instead of nnU-Net ResEnc L on KiTS | 3D segmentation · KiTS DSC | DSC: 88.98 -> 88.55 | Hurt |
| Use SonicNet instead of nnU-Net ResEnc L on ACDC | 3D segmentation · ACDC DSC | DSC: 91.40 -> 92.02 | Improved |
| Use SonicNet instead of nnU-Net with fewer parameters | 3D segmentation · parameters | parameters (M): 31.12 -> 2.59 | Improved |
| Use SonicNet instead of nnU-Net on PROMIS | external validation · PROMIS AUROC | AUROC: 0.646 -> 0.687 | Improved |
| Use SonicNet instead of nnU-Net on Prostate158 | external validation · Prostate158 AUROC | AUROC: 0.814 -> 0.841 | Improved |
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.
- S4ND — S4ND: Modeling Images and Videos as Multidimensional Signals with State Spaces (NeurIPS 2022) · better on 4 of 5 shared results
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
- Harmonic Networks: Deep Translation and Rotation Equivariance (CVPR 2017 · similarity 0.88)
- Fast Fourier Convolution (NeurIPS 2020 · similarity 0.86)
- Shape correspondence using anisotropic Chebyshev spectral CNNs (CVPR 2020 · similarity 0.86)
- Spectral State Space Model for Rotation-Invariant Visual Representation Learning (CVPR 2025 · similarity 0.86)
- HoloNets: Spectral Convolutions do extend to Directed Graphs (ICLR 2024 · similarity 0.85)
- Spectral Prompt Tuning: Unveiling Unseen Classes for Zero-Shot Semantic Segmentation (AAAI 2024 · 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.
