Research map / Architectures and efficiency / RNNs and normalization
One Neuron Saved is One Neuron Earned: On Parametric Efficiency of Quadratic Networks
Fenglei Fan, Hangcheng Dong, Zhongming Wu, Lecheng Ruan, Tieyong Zeng, Yiming Cui, Jing-Xiao Liao
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 |
|---|---|---|---|
| replace conventional neurons with quadratic neurons | one-hidden-layer network · Concentric hyperspheres (d=200) | accuracy (%): 82.75 -> 100.00 | Improved |
| replace conventional neurons with quadratic neurons | one-hidden-layer network · Concentric hyperspheres (d=100) | accuracy (%): 88.25 -> 100.00 | Improved |
| replace conventional neurons with quadratic neurons | one-hidden-layer network · Concentric hyperspheres (d=3) | accuracy (%): 100.00 -> 100.00 | No clear effect |
| use quadratic network with fewer neurons | one-hidden-layer network · Gaussian mixture (20-dim) | accuracy (%): 79.60 -> 82.00 | Improved |
| use quadratic network at matched parameter count | one-hidden-layer network · Gaussian mixture (500-dim) | accuracy (%): 66.60 -> 67.40 | Improved |
| add more neurons/depth to conventional network (ablation) | conventional network · Gaussian mixture (20-dim) | accuracy (%): 79.60 -> 79.20 | No clear effect |
| replace six layers in stage four of ResNet-50 with a 3-layer quadratic network | ResNet-50 · ImageNet-1K val | top-1 accuracy (%): 75.3 -> 75.4 | No clear effect |
| use quadratic aggregation instead of widening (ablation) | PointNet · S3DIS | mIOU: 45.2 -> 48.1 | Improved |
| use quadratic aggregation instead of stacking more layers (ablation) | PointNet · S3DIS | mIOU: 46.1 -> 48.1 | Improved |
| use quadratic aggregation instead of linear | DGCNN · S3DIS | mAcc: 57.0 -> 61.9 | Improved |
| use quadratic aggregation instead of widening (ablation) | DGCNN · S3DIS | mIOU: 49.6 -> 50.6 | Improved |
| make DGCNN deeper (ablation) | DGCNN · S3DIS | mIOU: 49.3 -> 49.9 | No clear effect |
Closest papers by meaning
- Dendritic Integration Inspired Artificial Neural Networks Capture Data Correlation (NeurIPS 2024 · similarity 0.86)
- The power of deeper networks for expressing natural functions (ICLR 2018 · similarity 0.86)
- Expand Neurons, Not Parameters (ICML 2026 · similarity 0.86)
- A Non-Local Convergence Analysis of Gradient Flow for Deep Linear Networks (TPAMI 2026 · similarity 0.85)
- The Effect of Weight Precision on the Neuron Count in Deep ReLU Networks (ICML 2024 · similarity 0.85)
- Learning quadratic neural networks in high dimensions: SGD dynamics and scaling laws (NeurIPS 2025 · similarity 0.85)
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.
