Research map / Optimization / Implicit bias and generalization
Deep Architecture Connectivity Matters for Its Convergence: A Fine-Grained Analysis
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 |
|---|---|---|---|
| Pre-filter architectures using effective depth and width | NAS-WOT · ImageNet | top-1 test error (%): 26.2 -> 25.9 (0.4) | Helped |
| Pre-filter architectures using effective depth and width | NAS-WOT · ImageNet | search cost (GPU days): 0.0036 -> 0.003 | Helped |
| Pre-filter architectures using effective depth and width | TE-NAS · ImageNet | top-1 test error (%): 24.5 -> 24.2 (0.3) | Helped |
| Pre-filter architectures using effective depth and width | TE-NAS · ImageNet | search cost (GPU days): 0.17 -> 0.1 | Helped |
| Pre-filter architectures using effective depth and width | WOT · NAS-Bench-201, CIFAR-10/CIFAR-100/ImageNet-16-120 | test accuracy (%): 92.81 (0.99) -> 92.98 (0.78); 69.48 (1.70) -> 69.86 (1.24); 43.10 (3.16) -> 43.44 (2.64) | Helped |
| Pre-filter architectures using effective depth and width | TE-NAS · NAS-Bench-201, CIFAR-10/CIFAR-100/ImageNet-16-120 | test accuracy (%): 93.9 (0.47) -> 93.6 (0.37); 71.24 (0.56) -> 71.26 (0.77); 42.38 (0.46) -> 44.38 (0.76) | Mixed |
| Pre-filter architectures using effective depth and width | FP-NAS · NAS-Bench-201 | search cost (GPU sec.): 4837 -> 2612 (-46.0%) | Helped |
| Restrict architectures to one half of the effective depth and width radii (ablation) | WOT + DAG · NAS-Bench-201 CIFAR-100 | test accuracy (%): 69.48 (1.70) -> 69.86 (1.24) | Helped |
| Restrict architectures to one quarter of the effective depth and width radii (ablation) | WOT + DAG · NAS-Bench-201 CIFAR-100 | test accuracy (%): 69.48 (1.70) -> 69.97 (1.19) | Helped |
| Use one half rather than three quarters of the effective depth and width radii (ablation) | WOT + DAG · NAS-Bench-201 CIFAR-100 | test accuracy (%): 69.51 (1.70) -> 69.86 (1.24) | No clear effect |
| Use the mixed sequential/parallel connectivity pattern (DAG#3) rather than parallel or sequential connections | Three-layer DAG network · MNIST and CIFAR-10 | training accuracy convergence speed | Helped |
Closest papers by meaning
- Understanding Mode Connectivity via Parameter Space Symmetry (ICML 2025 · similarity 0.87)
- Input Space Mode Connectivity in Deep Neural Networks (ICLR 2025 · similarity 0.87)
- On Linear Mode Connectivity of Mixture-of-Experts Architectures (NeurIPS 2025 · similarity 0.86)
- Topological Generalization Bounds for Discrete-Time Stochastic Optimization Algorithms (NeurIPS 2024 · similarity 0.86)
- Optimizer-Induced Mode Connectivity: From AdamW to Muon (NeurIPS 2026 · similarity 0.86)
- Flat Channels to Infinity in Neural Loss Landscapes (NeurIPS 2025 · similarity 0.85)
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.