Research map / Architectures and efficiency / CNN design
Scaling the Scattering Transform: Deep Hybrid Networks
Edouard Oyallon, Eugene Belilovsky, Sergey Zagoruyko
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 SIFT+Fisher-vector encoding with scattering + shared local encoder | FV + FC · ImageNet ILSVRC2012 val | top-1 accuracy: 55.6 -> 57.0 | Improved |
| replace SIFT+Fisher-vector (SVM) with scattering + SLE | FV + SVM · ImageNet ILSVRC2012 val | top-1 accuracy: 54.3 -> 57.0 | Improved |
| use scattering + SLE (6 learned layers) instead of AlexNet | AlexNet · ImageNet ILSVRC2012 val | top-1 error: 43.1 vs 43.0 | No clear effect |
| replace early ResNet layers with fixed scattering (Resnet-10 hybrid) | Resnet-18 · ImageNet ILSVRC2012 val | top-5 accuracy: 88.8 -> 88.6; 11.7M -> 12.8M params | No clear effect |
| add scattering front-end to wide ResNet in small-sample regime | WRN 16-8 · CIFAR-10 (100 training samples) | accuracy: 34.7 -> 38.9 | Improved |
| add scattering front-end to wide ResNet in small-sample regime | WRN 16-8 · CIFAR-10 (500 training samples) | accuracy: 46.5 -> 54.7 | Improved |
| add scattering front-end to wide ResNet in small-sample regime | WRN 16-8 · CIFAR-10 (1000 training samples) | accuracy: 60.0 -> 62.0 | Improved |
| use hybrid scattering network at full CIFAR-10 data instead of end-to-end WRN | Scat + Resnet · CIFAR-10 full | accuracy: 93.1 vs 96.0 | Hurt |
| add learned SLE layers F1, F2, F3 on top of scattering (ablation) | Raw scattering + SVM · Caltech-101 | accuracy: 62.8 -> 82.9 | Improved |
| use scattering hybrid on STL-10 without unlabeled data | Scat + WRN 19-8 · STL-10 | accuracy: 70.1 -> 76.0 | Improved |
Closest papers by meaning
- Scattering Networks for Hybrid Representation Learning (TPAMI 2018 · similarity 0.95)
- Compressing the Input for CNNs with the First-Order Scattering Transform (ECCV 2018 · similarity 0.90)
- Deep Network Classification by Scattering and Homotopy Dictionary Learning (ICLR 2020 · similarity 0.88)
- Parametric Scattering Networks (CVPR 2022 · similarity 0.87)
- Structured Receptive Fields in CNNs (CVPR 2016 · similarity 0.86)
- Unsupervised Parameter-free Simplicial Representation Learning with Scattering Transforms (ICML 2024 · similarity 0.85)
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.
