Research map / Architectures and efficiency / RNNs and normalization
Numerical influence of ReLU’(0) on backpropagation
David Bertoin, Jerome Bolte, Sébastien Gerchinovitz, Edouard Pauwels
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 |
|---|---|---|---|
| switch from 32 to 64 bit precision | fully connected net MNIST · MNIST weight sampling | proportion of weights in bifurcation zone: 40% -> 0% | Improved |
| switch from 32 to 16 bit precision | fully connected net MNIST · MNIST weight sampling | proportion of weights in bifurcation zone: 40% -> 100% | Hurt |
| set ReLU'(0)=1 instead of 0 | ResNet50 SGD · ImageNet | accuracy: 75% to 63% or 55% | Hurt |
| add batch normalization | VGG11 · CIFAR10 | test accuracy sensitivity to ReLU'(0) | Improved |
| use Adam instead of SGD | VGG11 · CIFAR10, SVHN | sensitivity to ReLU'(0) | Improved |
| use 64 bit precision with ReLU'(0)=1 | VGG11 · CIFAR10 | test performance | Improved |
| train with ReLU'(0)=1 in 16 bits | VGG11 · CIFAR10 | trainability | Hurt |
| vary ReLU'(0) magnitude beyond 0.1 without batch norm | VGG11 · CIFAR10 | test accuracy | Hurt |
| add dropout with varying probability | fully connected net · MNIST | interaction with ReLU'(0) | No clear effect |
| increase sample size | fully connected net · MNIST weight sampling | probability of bifurcation zone | Hurt |
| increase network width and depth | fully connected net · MNIST weight sampling | probability of bifurcation zone | Hurt |
| vary mini-batch size | fully connected net · MNIST weight sampling | probability of bifurcation zone | No clear effect |
Closest papers by meaning
- [Re] Numerical influence of ReLU'(0) on backpropagation (NeurIPS 2023 · similarity 0.89)
- The Effect of Weight Precision on the Neuron Count in Deep ReLU Networks (ICML 2024 · similarity 0.86)
- Why ReLU? A Bit-Model Dichotomy for Deep Network Training (ICML 2026 · similarity 0.86)
- Simplicity Bias and Optimization Threshold in Two-Layer ReLU Networks (ICML 2025 · similarity 0.84)
- Optimal Rates for Generalization of Gradient Descent for Deep ReLU Classification (NeurIPS 2025 · similarity 0.84)
- ReLUs Are Sufficient for Learning Implicit Neural Representations (ICML 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.
