Research map / Architectures and efficiency / RNNs and normalization
High-Capacity Expert Binary Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Add data-specific single-expert binary convolutions | Binary ResNet-18 · ImageNet | Top-1 accuracy: 60.9 -> 63.8 | Helped |
| Expand width using grouped convolutions (ablation) | Binary network with one expert · ImageNet | Top-1 accuracy: 60.9 -> 64.6 | Helped |
| Combine width expansion 2 with 8 experts (ablation) | Binary network · ImageNet | Top-1 accuracy: 60.9 -> 66.3 | Helped |
| Use stronger teacher for AT+KD (ablation) | Binary student · ImageNet | Top-1 accuracy: 70.0 -> 71.2 | Helped |
| Replace the gating softmax gradient with sigmoid (ablation) | EBConv · ImageNet Stage I | accuracy: 65.5% -> 62.7% | Hurt |
| Vary softmax temperature from 0.02 to 5 (ablation) | EBConv gating · ImageNet Stage I | Top-1 accuracy: 65.4 -> 65.5 -> 65.4 | No clear effect |
| Add Mixup to random scaling and cropping (ablation) | Proposed binary network · ImageNet | accuracy: +0.4% | Helped |
| Add Mixup (ablation) | Real-to-Bin · ImageNet Stage I and Stage II | accuracy: drops by 0.25% for Stage I, and 0.8% for Stage II | Hurt |
| Use CondConv's weighted expert combination instead of single-expert selection (ablation) | Binary ResNet-18 · ImageNet | Top-1 accuracy: 63.8% vs 61.2% | Hurt |
| Rearrange blocks across network stages (ablation) | Binary ResNet-18 Stage I · ImageNet | Top-1 accuracy: 63.7 to 63.9 | No clear effect |
| Use stronger data augmentation on Real-to-Bin (ablation) | Real-to-Bin · ImageNet | accuracy: drops by 0.25% for Stage I, and 0.8% for Stage II | Hurt |
| Use PReLU rather than linear downsampling decomposition (ablation) | Binary network · ImageNet Stage I | Top-1 accuracy: 66.5 -> 67.5 | Helped |
Closest papers by meaning
- Training binary neural networks with real-to-binary convolutions (ICLR 2020 · similarity 0.89)
- BD-Net: Has Depth-Wise Convolution Ever Been Applied in Binary Neural Networks? (AAAI 2026 · similarity 0.89)
- A Main/Subsidiary Network Framework for Simplifying Binary Neural Networks (CVPR 2019 · similarity 0.88)
- How to Train a Compact Binary Neural Network with High Accuracy? (AAAI 2017 · similarity 0.88)
- Training Binary Neural Networks via Gaussian Variational Inference and Low-Rank Semidefinite Programming (NeurIPS 2024 · similarity 0.85)
- Layerwise Progressive Freezing: A Training Scaffold for Depth-Scalable Binary Networks (NeurIPS 2026 · 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.