Research map / Architectures and efficiency / RNNs and normalization
Characterizing the Discrete Geometry of ReLU Networks
Blake B. Gaines, Jinbo Bi
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 |
|---|---|---|---|
| increase depth from 1 to 4 at width 4 (ablation) | ReLU network · Synthetic Gaussians, input dim 4 | average degree: 4.00±0.00 -> 6.17±0.12 | Improved |
| increase width from 4 to 16 at depth 1 (ablation) | ReLU network · Synthetic Gaussians, input dim 4 | average degree: 4.00±0.00 -> 7.32±0.00 | Improved |
| increase depth from 1 to 4 at width 16 (ablation) | ReLU network · Synthetic Gaussians, input dim 4 | average degree: 7.32±0.00 -> 7.85±0.03 | Improved |
| increase input dimension from 4 to 5 at depth 4, width 16 (ablation) | ReLU network · Synthetic Gaussians | average degree: 7.85±0.03 -> 9.80±0.03 | Improved |
| increase depth from 1 to 4 at width 16 (ablation) | ReLU network · Synthetic Gaussians, input dim 4 | graph diameter: 22.50±0.35 -> 76.35±4.56 | Hurt |
| increase width from 4 to 16 at depth 1 (ablation) | ReLU network · Synthetic Gaussians, input dim 4 | graph diameter: 5.50±0.00 -> 22.50±0.35 | Hurt |
| change input dimension from 4 to 5 at depth 1, width 8 (ablation) | ReLU network · Synthetic Gaussians | graph diameter: 10.60±0.42 -> 10.75±0.26 | No clear effect |
| change input dimension from 4 to 5 at depth 1, width 16 (ablation) | ReLU network · Synthetic Gaussians | graph diameter: 22.50±0.35 -> 23.17±0.41 | No clear effect |
| compare regions containing training data vs regions without data | trained ReLU networks · MNIST, CIFAR10, California Housing | neighbor count | Improved |
| compare proportion of unbounded data-containing regions, classification vs regression | trained ReLU networks · MNIST, CIFAR10 vs California Housing | fraction of unbounded regions | Mixed |
Closest papers by meaning
- The Computational Complexity of Counting Linear Regions in ReLU Neural Networks (NeurIPS 2025 · similarity 0.88)
- Depth-Bounds for Neural Networks via the Braid Arrangement (NeurIPS 2025 · similarity 0.87)
- The Symmetries of Three-Layer ReLU Networks (NeurIPS 2026 · similarity 0.86)
- On the Local Complexity of Linear Regions in Deep ReLU Networks (ICML 2025 · similarity 0.86)
- Most ReLU Networks Admit Identifiable Parameters (NeurIPS 2026 · similarity 0.86)
- Topological obstruction to the training of shallow ReLU neural networks (NeurIPS 2024 · similarity 0.86)
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.
