Research map / Architectures and efficiency / RNNs and normalization
Truncating Wide Networks Using Binary Tree Architectures
Yan Zhang, Mete Ozay, Shuohao Li, Takayuki Okatani
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 Wide ResNet blocks with binary tree blocks (d=12,k=2,n=4) | Wide ResNet · CIFAR-100 | classification error (%): 20.43 -> 19.06 | Improved |
| replace Wide ResNet blocks with binary tree blocks (d=12,k=4,n=2) | Wide ResNet · CIFAR-100 | classification error (%): 20.43 -> 19.22 | Improved |
| replace Wide ResNet blocks with binary tree blocks (d=10,k=2,n=3) | Wide ResNet · CIFAR-100 | classification error (%): 20.75 -> 19.29 | Improved |
| replace Wide ResNet blocks with binary tree blocks (d=10,k=2,n=2) | Wide ResNet · CIFAR-100 | classification error (%): 21.59 -> 20.48 | Improved |
| replace Wide ResNet blocks with binary tree blocks (d=4,k=3,n=4) | Wide ResNet · CIFAR-100 | classification error (%): 22.89 -> 22.19 | Improved |
| use a single BitBlock per group with k=4 (d=10,k=4,n=1) | BitNet · CIFAR-100 | classification error (%): 21.59 -> 23.88 | Hurt |
| replace Wide ResNet blocks with binary tree blocks (d=4,k=6,n=2) | Wide ResNet · CIFAR-100 | classification error (%): 22.89 -> 23.87 | Hurt |
| add binary tree block depth k from 1 to 2 (d=12,n=2) (ablation) | BitNet · CIFAR-100 | classification error (%): 23.80 -> 19.86 | Improved |
| increase block depth k beyond 3 (d=12,n=2) (ablation) | BitNet · CIFAR-100 | classification error (%): 18.97 -> 19.22 | No clear effect |
| add binary tree block depth k from 1 to 3 (d=4,n=4) (ablation) | BitNet · CIFAR-100 | classification error (%): 23.54 -> 22.19 | Improved |
| add concatenation-based BitBlock (b(4,4,3)-38) versus plain network without shortcuts (ablation) | plainnet-38 · CIFAR-100 | classification error (%): 29.13 -> 22.60 | Improved |
| increase number of blocks n and reduce k (b(4,2,6)-38 versus b(4,6,2)-38) (ablation) | BitNet depth 38 · CIFAR-100 | first-layer gradient magnitude | Hurt |
Closest papers by meaning
- BCNet: Searching for Network Width With Bilaterally Coupled Network (CVPR 2021 · similarity 0.86)
- Searching for Network Width With Bilaterally Coupled Network (TPAMI 2022 · similarity 0.86)
- When narrower is better: the narrow width limit of Bayesian parallel branching neural networks (ICLR 2025 · similarity 0.85)
- Depthwise Hyperparameter Transfer in Residual Networks: Dynamics and Scaling Limit (ICLR 2024 · similarity 0.84)
- Unveiling The Matthew Effect Across Channels: Assessing Layer Width Sufficiency via Weight Norm Variance (NeurIPS 2024 · similarity 0.84)
- How Sparse Can We Prune A Deep Network: A Fundamental Limit Perspective (NeurIPS 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.
