Research map / Architectures and efficiency / RNNs and normalization
How to Initialize your Network? Robust Initialization for WeightNorm & ResNets
Devansh Arpit, Víctor Campos, Yoshua Bengio
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 the proposed initialization instead of data-dependent initialization | Weight-normalized ResNet-56 · CIFAR-10 | test error (%): 9.19 $\pm$ 0.24 -> 7.87 $\pm$ 0.14 | Improved |
| Add learning-rate warmup to proposed initialization (ablation) | Weight-normalized ResNet-56 · CIFAR-10 | test error (%): 7.87 $\pm$ 0.14 -> 7.20 $\pm$ 0.12 | Improved |
| Use the proposed initialization instead of data-dependent initialization | Weight-normalized ResNet-110 · CIFAR-10 | test error (%): 9.33 $\pm$ 0.10 -> 7.71 $\pm$ 0.14 | Improved |
| Add learning-rate warmup to proposed initialization (ablation) | Weight-normalized ResNet-110 · CIFAR-10 | test error (%): 7.71 $\pm$ 0.14 -> 6.69 $\pm$ 0.11 | Improved |
| Use the proposed initialization instead of data-dependent initialization | Weight-normalized WRN-40-10 with cutout · CIFAR-10 | test error (%): 6.10 $\pm$ 0.23 -> 4.74 $\pm$ 0.14 | Improved |
| Add learning-rate warmup to proposed initialization with cutout (ablation) | Weight-normalized WRN-40-10 · CIFAR-10 | test error (%): 4.74 $\pm$ 0.14 -> 4.75 $\pm$ 0.08 | No clear effect |
| Use the proposed initialization instead of data-dependent initialization | Weight-normalized ResNet-164 with cutout · CIFAR-100 | test error (%): 30.26 $\pm$ 0.51 -> 27.30 $\pm$ 0.49 | Improved |
| Add learning-rate warmup to proposed initialization with cutout (ablation) | Weight-normalized ResNet-164 · CIFAR-100 | test error (%): 27.30 $\pm$ 0.49 -> 25.31 $\pm$ 0.26 | Improved |
| Use the proposed initialization instead of data-dependent initialization | Weight-normalized residual networks · CIFAR-10 and CIFAR-100; test error | test error (%): 9.19 $\pm$ 0.24 -> 7.87 $\pm$ 0.14; 9.33 $\pm$ 0.10 -> 7.71 $\pm$ 0.14; 6.10 $\pm$ 0.23 -> 4.74 $\pm$ 0.14; 30.26 $\pm$ | Improved |
| Use proposed initialization with warmup instead of data-dependent initialization | Weight-normalized residual networks · CIFAR-10 ResNet-56 and ResNet-110; CIFAR-100 ResNet-164 | test error (%): 9.19 $\pm$ 0.24 -> 7.20 $\pm$ 0.12; 9.33 $\pm$ 0.10 -> 6.69 $\pm$ 0.11; 30.26 $\pm$ 0.51 -> 25.31 $\pm$ 0.26 | Improved |
| Use the proposed initialization rather than PyTorch default | Weight-normalized residual networks · CIFAR-10 WRNs across evaluated depths | successful training across depths | Improved |
| Use the proposed initialization rather than data-dependent initialization | Weight-normalized fully connected networks · MNIST, depths up to 200 layers | trainability at depth | Improved |
Closest papers by meaning
- Principled Weight Initialization for Hypernetworks (ICLR 2020 · similarity 0.88)
- Initializing Models with Larger Ones (ICLR 2024 · similarity 0.85)
- Sinusoidal Initialization, Time for a New Start (NeurIPS 2025 · similarity 0.85)
- If You Want to Be Robust, Be Wary of Initialization (NeurIPS 2024 · similarity 0.85)
- Robust Weight Initialization for Tanh Neural Networks with Fixed Point Analysis (ICLR 2025 · similarity 0.84)
- IDInit: A Universal and Stable Initialization Method for Neural Network Training (ICLR 2025 · 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.
