Research map / Architectures and efficiency / Neural architecture search
Discovering Neural Wirings
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 |
|---|---|---|---|
| Learn edge connectivity during training rather than fixing a random wiring | Static neural graph on CIFAR-10 · CIFAR-10 | accuracy: 76.1 ± 0.5% -> 80.9 ± 0.6% | Helped |
| Learn edge connectivity during training rather than fixing a random wiring | Discrete-time neural graph on CIFAR-10 · CIFAR-10 | accuracy: 77.3 ± 0.7% -> 82.3 ± 0.6% | Helped |
| Learn edge connectivity during training rather than fixing a random wiring | Continuous-time neural graph on CIFAR-10 · CIFAR-10 | accuracy: 78.5 ± 1.2 % -> 83.1 ± 0.3 % | Helped |
| Jointly learn MobileNetV1 channel connectivity instead of using its hand-designed wiring | MobileNetV1-DNW ×0.225 · ImageNet | accuracy: 50.6% -> 60.9% | Helped |
| Learn wiring instead of fixing a random graph | MobileNetV1-derived network at similar compute · ImageNet | accuracy: 53.3% -> 60.9% | Helped |
| Learn a sparse subnetwork within ResNet50 | ResNet50 with 10% weights · ImageNet | Top-1 accuracy: 72.9% -> 74.0% | Helped |
| Keep the first convolution dense (ablation) | DNW ResNet50 with 10% weights · ImageNet | Top-1 accuracy: 74.0% -> 75.0% | Helped |
| Keep the first convolution dense (ablation) | DNW ResNet50 with 20% weights · ImageNet | Top-1 accuracy: 76.2% -> 76.6% | Helped |
| Keep the first convolution dense (ablation) | DNW ResNet50 with 30% weights · ImageNet | Top-1 accuracy: 76.9% -> 77.1% | Helped |
| Remove the update rule for hallucinated edges (ablation) | MobileNetV1-DNW architecture · CIFAR-10 | accuracy: 89.7 ± 0.2% -> 86.7 ± 0.5% | Hurt |
| Use unconstrained targeted dropout across edges at a resolution instead of regular per-neuron dropout (ablation) | Targeted Dropout ρ=0.95 · CIFAR-10 | accuracy: 87.9 ± 0.2% -> 89.2 ± 0.4% | Helped |
| Vary targeted dropout probability (ablation) | Unconstrained Targeted Dropout · CIFAR-10 | accuracy: 89.0 ± 0.2% (ρ=0.9); 89.2 ± 0.4% (ρ=0.95); 88.6 ± 0.2% (ρ=0.99); 88.8 ± 0.2% (ρ=γ) | Mixed |
Closest papers by meaning
- Deep Architecture Connectivity Matters for Its Convergence: A Fine-Grained Analysis (NeurIPS 2022 · similarity 0.89)
- Neural Architecture Search as Sparse Supernet (AAAI 2021 · similarity 0.88)
- SMASH: One-Shot Model Architecture Search through HyperNetworks (ICLR 2018 · similarity 0.88)
- ISTA-NAS: Efficient and Consistent Neural Architecture Search by Sparse Coding (NeurIPS 2020 · similarity 0.88)
- Neural Architecture Search on ImageNet in Four GPU Hours: A Theoretically Inspired Perspective (ICLR 2021 · similarity 0.88)
- Efficient Few-Shot Neural Architecture Search by Counting the Number of Nonlinear Functions (AAAI 2025 · similarity 0.87)
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.