Research map / Architectures and efficiency / RNNs and normalization
Network Morphism
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 |
|---|---|---|---|
| NetMorph instead of Net2Net for MNIST softmax to MLP (PReLU) (ablation) | network morphism · MNIST | accuracy: 92% -> 97% vs 94 | Helped |
| NetMorph with non-idempotent PTanH activation (ablation) | NetMorph · MNIST | accuracy | Helped |
| NetMorph vs Net2Net in depth morphing cifar_111->211->222 (ablation) | network morphism · CIFAR10 | accuracy: 82.06% vs 81.21%; 82.43% vs 81.99 | Helped |
| subnet morphing cifar_222 to cifar_2222 | NetMorph · CIFAR10 | accuracy: 82.43% -> 83.14 | Helped |
| subnet morphing cifar_2222 to cifar_3333 | NetMorph · CIFAR10 | accuracy: 83.14% -> 84 | Helped |
| train NetMorph in phases versus training the deep net from scratch (Raw) (ablation) | NetMorph · CIFAR10 | accuracy | Helped |
| kernel size morphing from 1 to 3 in second layer of each subnet | NetMorph · CIFAR10 | accuracy: 81.48% -> 82.81 | Helped |
| width morphing, double channels of first layer per subnet: NetMorph vs Net2Net (ablation) | network morphism · CIFAR10 | accuracy: 83.09% vs 82.70 | Helped |
| width morphing directly from parent with TanH neurons | NetMorph · CIFAR10 | accuracy: about 4 | Helped |
| morph VGG16 baseline by adding conv layers to first three subnets | NetMorph · ImageNet | top-1 1-view accuracy: 67.30% -> 69.14 | Helped |
| morph VGG16 vs VGG19 multi-scale | NetMorph · ImageNet | top-1 10-view accuracy: 69.44% -> 70.32 | Helped |
| NetMorph training schedule vs training VGG from scratch | NetMorph · ImageNet | training time: 2~3 months -> 5 days (15x) | Helped |
Closest papers by meaning
- MorphNet: Fast & Simple Resource-Constrained Structure Learning of Deep Networks (CVPR 2018 · similarity 0.88)
- On the Hardness of Training Deep Neural Networks Discretely (AAAI 2025 · similarity 0.85)
- Mitigating Over-squashing without Rewiring: A Sheaf Effective Resistance Perspective (NeurIPS 2026 · similarity 0.85)
- Expand-and-Cluster: Parameter Recovery of Neural Networks (ICML 2024 · similarity 0.84)
- On Vanishing Gradients, Over-Smoothing, and Over-Squashing in GNNs: Bridging Recurrent and Graph Learning (NeurIPS 2025 · similarity 0.84)
- Making Classic GNNs Strong Baselines Across Varying Homophily: A Smoothness–Generalization Perspective (NeurIPS 2025 · 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.