Research map / Architectures and efficiency / RNNs and normalization
Neural Network Architecture Beyond Width and Depth
Shijun Zhang, Zuowei Shen, Haizhao Yang
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 |
|---|---|---|---|
| Use a trainable sub-network activation function (varrho) instead of ReLU | NestNet versus standard network, width 20 and depth 4 · Archimedean spiral test | test accuracy: 0.738290 -> 0.873631 | Improved |
| Use a trainable sub-network activation function (varrho) instead of ReLU | NestNet versus standard network, width 35 and depth 4 · Archimedean spiral test | test accuracy: 0.816048 -> 0.995962 | Improved |
| Use a trainable sub-network activation function (varrho) instead of ReLU | NestNet versus standard network, width 50 and depth 4 · Archimedean spiral test | test accuracy: 0.866118 -> 0.999984 | Improved |
| Use sub-network activation varrho for a few CNN activations instead of ReLU | CNN2 versus CNN1 · Fashion-MNIST test | average accuracy in last 100 epochs: 0.924447 -> 0.926032 | Improved |
| Use sub-network activation varrho for a few CNN activations instead of ReLU | CNN2 versus CNN1 · Fashion-MNIST training | training time: ≈ 5802 s -> ≈ 7217 s | Hurt |
| Add nested architecture height s while keeping parameter count O(n) | ReLU networks approximating 1-Lipschitz functions on [0,1]^d · 1-Lipschitz continuous functions on [0,1]^d | approximation error: n^{-2/d} -> n^{-(s+1)/d} | Improved |
Closest papers by meaning
- The Expressive Power of Neural Networks: A View from the Width (NeurIPS 2017 · similarity 0.88)
- Quasi-Equivalence between Width and Depth of Neural Networks (JMLR 2023 · similarity 0.86)
- Neural Networks with Small Weights and Depth-Separation Barriers (NeurIPS 2020 · similarity 0.86)
- On the Expressiveness of Rational ReLU Neural Networks With Bounded Depth (ICLR 2025 · similarity 0.85)
- Compelling ReLU Networks to Exhibit Exponentially Many Linear Regions at Initialization and During Training (ICML 2025 · similarity 0.85)
- ReLU Network with Width $d+\mathcal{O}(1)$ Can Achieve Optimal Approximation Rate (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.
