Research map / Architectures and efficiency / CNN design
Scattering Networks for Hybrid Representation Learning
Edouard Oyallon, Sergey Zagoruyko, Gabriel Huang, Nikos Komodakis, Simon Lacoste-Julien, Matthew B. Blaschko, Eugene Belilovsky
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 early learned layers with a scattering network plus 1x1 shared local encoder | AlexNet-like pipeline · ILSVRC2012 val | top-1 accuracy (%): 56.9 -> 57.0 | No clear effect |
| replace early layers of ResNet with scattering, learning at 28x28 resolution | Resnet-18 · ILSVRC2012 val | top-5 accuracy (%): 88.8 -> 88.6 | No clear effect |
| add scattering front end to wide ResNet in 100-sample regime | WRN 16-8 · CIFAR-10 (100 samples) | accuracy (%): 34.7 -> 38.9 | Improved |
| add scattering front end to wide ResNet in 500-sample regime | WRN 16-8 · CIFAR-10 (500 samples) | accuracy (%): 46.5 -> 54.7 | Improved |
| add scattering front end to wide ResNet in 1000-sample regime | WRN 16-8 · CIFAR-10 (1000 samples) | accuracy (%): 60.0 -> 62.0 | Improved |
| add scattering front end to wide ResNet with full training data | WRN 16-8 · CIFAR-10 (full) | accuracy (%): 95.7 -> 93.1 | Hurt |
| use scattering hybrid instead of supervised CNN on STL-10 without unlabeled data | CNN · STL-10 | accuracy (%): 70.1 -> 76.0 | Improved |
| use scattering representation instead of random CNN features with linear classifier | random CNN · ImageNet linear classification | top-1 accuracy (%): 12.9 -> 17.4 | Improved |
| add successive SLE layers F1, F2, F3 on raw scattering for Caltech-101 | Scattering + linear SVM · Caltech-101 | accuracy (%): 62.8 -> 82.9 | Improved |
| use scattering + FC unsupervised representation instead of DCGAN features | DCGAN · CIFAR-10 (unsupervised) | accuracy (%): 82.8 -> 84.7 | Improved |
| replace scattering by a learned 5-layer CNN approximator (ablation) | Scat + Cascaded CNN · CIFAR-10 | accuracy (%): 91.4 -> about 1% lower | Hurt |
Closest papers by meaning
- Scaling the Scattering Transform: Deep Hybrid Networks (ICCV 2017 · similarity 0.95)
- Compressing the Input for CNNs with the First-Order Scattering Transform (ECCV 2018 · similarity 0.91)
- Deep Network Classification by Scattering and Homotopy Dictionary Learning (ICLR 2020 · similarity 0.90)
- Structured Receptive Fields in CNNs (CVPR 2016 · similarity 0.88)
- Unsupervised Parameter-free Simplicial Representation Learning with Scattering Transforms (ICML 2024 · similarity 0.85)
- Exploitation of a Latent Mechanism in Graph Contrastive Learning: Representation Scattering (NeurIPS 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.
