Research map / Architectures and efficiency / CNN design
Separation and Concentration in Deep Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 learned channel projectors to the fixed scattering tree (ablation) | Scattering network · CIFAR-10 | classification error (%): 27.7 -> 12.8 | Helped |
| Add learned channel projectors to the fixed scattering tree (ablation) | Scattering network · ImageNet validation | Top-1 error (%): 73.0 -> 42.3 | Helped |
| Add tight-frame thresholding concentration after learned scattering projections (ablation) | S_P · CIFAR-10 | classification error (%): 12.8 -> 8.0 | Helped |
| Add tight-frame thresholding concentration after learned scattering projections (ablation) | S_P · ImageNet validation | Top-1 error (%): 42.3 -> 31.4 | Helped |
| Use thresholded ReLU instead of soft-thresholding (ablation) | S_C · CIFAR-10 | classification error (%): 8.0 -> 7.6 | Helped |
| Use thresholded ReLU instead of soft-thresholding (ablation) | S_C · ImageNet validation | Top-1 error (%): 31.4 -> 29.7 | Helped |
| Use thresholded ReLU instead of soft-thresholding (ablation) | S_C · ImageNet validation | Top-5 error (%): 11.6 -> 10.7 | Helped |
| Replace the input with a learned tight-frame nonlinearity using absolute value (ablation) | Logistic classifier · CIFAR-10 | classification error (%): 60.5 -> 28.1 | Helped |
| Use soft-thresholding rather than absolute value (ablation) | F^T rho F two-layer representation · CIFAR-10 | classification error (%): 28.1 -> 34.8 | Hurt |
| Use thresholded ReLU rather than soft-thresholding (ablation) | F^T rho F two-layer representation · CIFAR-10 | classification error (%): 34.8 -> 26.5 | Helped |
| Use a tight-frame nonlinearity rather than the raw input (ablation) | Logistic classifier · MNIST | classification error (%): 7.4 -> 1.3 | Helped |
| Use thresholded ReLU rather than soft-thresholding (ablation) | F^T rho F two-layer representation · MNIST | classification error (%): 1.4 -> 1.3 | No clear effect |
Closest papers by meaning
- Deep Network Classification by Scattering and Homotopy Dictionary Learning (ICLR 2020 · similarity 0.88)
- Maximum Class Separation as Inductive Bias in One Matrix (NeurIPS 2022 · similarity 0.88)
- Phase Collapse in Neural Networks (ICLR 2022 · similarity 0.85)
- Inductive Bias of Deep Convolutional Networks through Pooling Geometry (ICLR 2017 · similarity 0.85)
- Separation Power of Equivariant Neural Networks (ICLR 2025 · similarity 0.84)
- Few-shot Learner Parameterization by Diffusion Time-steps (CVPR 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.