Research map / Architectures and efficiency / RNNs and normalization
Parameter Prediction for Unseen Deep Architectures
Boris Knyazev, Michal Drozdzal, Graham W. Taylor, Adriana Romero
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 |
|---|---|---|---|
| replace GatedGNN with an MLP (ablation) | GHN-2 · CIFAR-10 ID-test | accuracy: 66.9 -> 42.2 | Hurt |
| remove operation-dependent normalization (ablation) | GHN-2 · CIFAR-10 ID-test | accuracy: 66.9 -> 62.6 | Hurt |
| remove normalization (ablation) | GHN-2 · CIFAR-10 OOD-test | accuracy: 56.8 -> 47.1 | Hurt |
| remove virtual edges (ablation) | GHN-2 · CIFAR-10 ID-test | accuracy: 66.9 -> 61.5 | Hurt |
| remove virtual edges (ablation) | GHN-2 · CIFAR-10 OOD-test | accuracy: 56.8 -> 53.9 | Hurt |
| remove meta-batching (b_m=1) (ablation) | GHN-2 · CIFAR-10 ID-test | accuracy: 66.9 -> 54.3 | Hurt |
| train with b_m=1 eight times longer (ablation) | GHN-2 · CIFAR-10 ID-test | accuracy: 66.9 -> 62.4 | Hurt |
| train on 1000 architectures instead of full set (ablation) | GHN-2 · CIFAR-10 ID-test | accuracy: 66.9 -> 65.1 | Hurt |
| GHN-2 versus GHN-1 on BN-free networks | GHN-2 · CIFAR-10 BN-free | accuracy: 13.7 -> 36.8 | Improved |
| GHN-2 versus GHN-1 on ID-test | GHN-2 · CIFAR-10 ID-test | accuracy: 51.4 -> 66.9 | Improved |
| GHN-2 versus GHN-1 on ImageNet ID-test | GHN-2 · ImageNet ID-test | top-5 accuracy: 17.2 -> 27.2 | Improved |
| GHN-2 initialization versus He init for ResNet-50 | ResNet-50 · 100-shot CIFAR-10 | accuracy: 41.0 -> 56.4 | Improved |
Closest papers by meaning
- Principled Architecture-aware Scaling of Hyperparameters (ICLR 2024 · similarity 0.89)
- Can We Scale Transformers to Predict Parameters of Diverse ImageNet Models? (ICML 2023 · similarity 0.88)
- TAPAS: Train-Less Accuracy Predictor for Architecture Search (AAAI 2019 · similarity 0.87)
- Bag of Tricks for Training Deeper Graph Neural Networks: A Comprehensive Benchmark Study (TPAMI 2022 · similarity 0.86)
- Depthwise Hyperparameter Transfer in Residual Networks: Dynamics and Scaling Limit (ICLR 2024 · similarity 0.86)
- Graph Metanetworks for Processing Diverse Neural Architectures (ICLR 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.
