Research map / Architectures and efficiency / RNNs and normalization
Meta-GF: Training Dynamic-Depth Neural Networks Harmoniously
Yi Sun, Jian Li, Xin Xu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| train with Meta-GF | MSDnet · CIFAR-100 | average top-1 accuracy over exits: 72.83 -> 74.57 | Improved |
| train with Meta-GF | MSDnet · CIFAR-10 | average top-1 accuracy over exits: 93.80 -> 94.54 | Improved |
| train with Meta-GF | MSDnet · ImageNet | average top-1 accuracy over exits: 66.72 -> 67.25 | Improved |
| train with Meta-GF | SDN-VGG · CIFAR-100 | average top-1 accuracy over exits: 66.34 -> 68.51 | Improved |
| train with Meta-GF | SDN-ResNet · CIFAR-100 | average top-1 accuracy over exits: 59.29 -> 60.32 | Improved |
| train with Meta-GF instead of Cagrad | SDN-ResNet · CIFAR-10 | average top-1 accuracy over exits: 87.04 -> 86.62 | Hurt |
| train with Meta-GF vs Cagrad | SDN-VGG · CIFAR-10 | average top-1 accuracy over exits: 89.54 -> 89.82 | Improved |
| train with Meta-GF instead of Pcgrad | MSDnet · CIFAR-100 | average top-1 accuracy over exits: 74.14 -> 74.57 | Improved |
| add Meta-GF on shallow exit | MSDnet · ImageNet exit 1 | average top-1 accuracy over exits: 58.48 -> 57.43 | Hurt |
| add Meta-GF on deepest exit | MSDnet · ImageNet exit 5 | top-1 accuracy: 71.03 -> 73.27 | Improved |
| add knowledge distillation to Meta-GF | MSDnet · CIFAR-100 (150 per class) | average top-1 accuracy over exits: 57.93 -> 59.87 | Improved |
| train with Meta-GF vs baseline | MSDnet · CIFAR-100 (150 per class) | average top-1 accuracy over exits: 56.38 -> 57.93 | Improved |
Closest papers by meaning
- Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks (ICML 2017 · similarity 0.86)
- Supplementary Meta-Learning: Towards a Dynamic Model for Deep Neural Networks (ICCV 2017 · similarity 0.85)
- Dynamic Kernel Selection for Improved Generalization and Memory Efficiency in Meta-Learning (CVPR 2022 · similarity 0.85)
- Adaptive Depth Networks with Skippable Sub-Paths (NeurIPS 2024 · similarity 0.83)
- MetaDiff: Meta-Learning with Conditional Diffusion for Few-Shot Learning (AAAI 2024 · similarity 0.83)
- Graph Metanetworks for Processing Diverse Neural Architectures (ICLR 2024 · similarity 0.83)
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.
