Research map / Architectures and efficiency / Pruning and efficient tuning
What's Hidden in a Randomly Weighted Neural Network?
Vivek Ramanujan, Mitchell Wortsman, Aniruddha Kembhavi, Ali Farhadi, Mohammad Rastegari
Check 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 |
|---|---|---|---|
| Use Signed Kaiming Constant rather than Kaiming Normal initialization (ablation) | edge-popup · ImageNet | top-1 accuracy: 61.71% -> 68.6% (ResNet-50); 66.15% -> 72.3% (ResNet-101); 67.95% -> 73.3% (Wide ResNet-50) | Improved |
| Increase parent network width and depth | edge-popup subnetworks with randomly weighted parent networks · CIFAR-10 and ImageNet | accuracy | Improved |
| Increase parent network width while holding subnetwork parameter count fixed (ablation) | edge-popup · CIFAR-10 | accuracy | Improved |
| Choose a middle sparsity level rather than a value near 0% or 100% (ablation) | edge-popup · CIFAR-10 | accuracy | Improved |
| Use a scaled initialization with the additional sqrt(1/k) factor (ablation) | edge-popup · ImageNet | top-1 accuracy | Improved |
| Use a scaled initialization with the additional sqrt(1/k) factor (ablation) | edge-popup · CIFAR-10 | accuracy | No clear effect |
| Train the weights of the discovered subnetwork after finding it (ablation) | edge-popup subnetworks · ImageNet | top-1 accuracy: 73.5% vs 73.6% (ResNet-50, Kaiming Normal); 76.5% vs 76.8% (Wide ResNet-50, Kaiming Normal); 76.1% vs 76.1% (ResNet-101, | No clear effect |
| Replace deterministic top-score edge selection with stochastic mask sampling | edge-popup versus Zhou et al.'s supermask method · CIFAR-10 and ImageNet | accuracy | Hurt |
Closest papers by meaning
- Peek-a-Boo: What (More) is Disguised in a Randomly Weighted Neural Network, and How to Find It Efficiently (ICLR 2022 · similarity 0.87)
- How Powerful are Shallow Neural Networks with Bandlimited Random Weights? (ICML 2023 · similarity 0.85)
- On the Power and Limitations of Random Features for Understanding Neural Networks (NeurIPS 2019 · similarity 0.84)
- How Uniform Random Weights Induce Non-uniform Bias: Typical Interpolating Neural Networks Generalize with Narrow Teachers (ICML 2024 · similarity 0.83)
- Neural Redshift: Random Networks are not Random Functions (CVPR 2024 · similarity 0.82)
- Expand-and-Cluster: Parameter Recovery of Neural Networks (ICML 2024 · similarity 0.82)
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.
