Research map / Architectures and efficiency / CNN design
Phase Collapse in Neural 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 1×1 channel projections and wavelet phase collapses instead of using the fixed scattering transform | Scattering network · CIFAR-10 | Top-1 error (%): 27.7 -> 11.7 | Helped |
| Add skip connections that concatenate wavelet coefficients with their moduli (ablation) | LScat · CIFAR-10 | Top-1 error (%): 11.7 -> 7.7 | Helped |
| Add skip connections that concatenate wavelet coefficients with their moduli (ablation) | LScat · ImageNet | Top-1 error (%): 35.9 -> 30.1 | Helped |
| Replace phase-collapse modulus with amplitude soft-thresholding (ablation) | LScat · CIFAR-10 | Top-1 error (%): 11.7 -> 36.7 | Hurt |
| Replace phase-collapse modulus with amplitude hyperbolic tangent (ablation) | LScat · CIFAR-10 | Top-1 error (%): 11.7 -> 40.7 | Hurt |
| Replace phase-collapse modulus with amplitude sigmoid (ablation) | LScat · CIFAR-10 | Top-1 error (%): 11.7 -> 38.5 | Hurt |
| Replace phase-collapse modulus with amplitude soft-sign (ablation) | LScat · CIFAR-10 | Top-1 error (%): 11.7 -> 39.9 | Hurt |
| Replace phase-collapse modulus with amplitude soft-thresholding (ablation) | LScat + skip · CIFAR-10 | Top-1 error (%): 7.7 -> 22.5 | Hurt |
| Replace phase-collapse modulus with amplitude-preserving-phase tanh (ablation) | LScat + skip · CIFAR-10 | Top-1 error (%): 7.7 -> 19.2 | Hurt |
| Replace phase-collapse modulus with amplitude sigmoid (ablation) | LScat + skip · CIFAR-10 | Top-1 error (%): 7.7 -> 17.0 | Hurt |
| Replace phase-collapse modulus with amplitude soft-sign (ablation) | LScat + skip · CIFAR-10 | Top-1 error (%): 7.7 -> 19.5 | Hurt |
| Use sign-preserving soft-thresholding instead of absolute-value sign collapse (ablation) | Bias-free ResNet-18 · ImageNet | Top-1 error (%): 35.3 -> 50.0 | Hurt |
Closest papers by meaning
- Neural Collapse under Gradient Flow on Shallow ReLU Networks for Orthogonally Separable Data (NeurIPS 2025 · similarity 0.87)
- Generalized Neural Collapse for a Large Number of Classes (ICML 2024 · similarity 0.86)
- Neural Collapse is Globally Optimal in Deep Regularized ResNets and Transformers (NeurIPS 2025 · similarity 0.86)
- Wide Neural Networks Trained with Weight Decay Provably Exhibit Neural Collapse (ICLR 2025 · similarity 0.85)
- Neural Collapse for Cross-entropy Class-Imbalanced Learning with Unconstrained ReLU Features Model (ICML 2024 · similarity 0.84)
- Neural collapse vs. low-rank bias: Is deep neural collapse really optimal? (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.