Research map / Architectures and efficiency / Quantization
Bi-Real Net: Binarizing Deep Network Towards Real-Network Performance
Zechun Liu, Wenhan Luo, Baoyuan Wu, Xin Yang, Wei Liu, Kwang-Ting Tim Cheng
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 |
|---|---|---|---|
| add real-valued shortcuts per 1-bit conv layer (Bi-Real structure vs plain) (ablation) | 18-layer 1-bit CNN · ImageNet val | top-1 accuracy: 12.1 -> 56.4 | Improved |
| 1-layer-per-block Bi-Real shortcut instead of 2-layer-per-block ResNet (ablation) | 18-layer 1-bit CNN · ImageNet val | top-1 accuracy: 45.7 -> 56.4 | Improved |
| add proposed activation derivative approximation (ablation) | Bi-Real-18 (ReLU init, original weight update) · ImageNet val | top-1 accuracy: 32.9 -> 36.8 | Improved |
| add magnitude-aware weight binarization (ablation) | Bi-Real-18 (ReLU init) · ImageNet val | top-1 accuracy: 32.9 -> 40.5 | Improved |
| replace ReLU with Clip initialization (ablation) | Bi-Real-18 · ImageNet val | top-1 accuracy: 32.9 -> 37.4 | Improved |
| add all three proposed training techniques (ablation) | Bi-Real-18 · ImageNet val | top-1 accuracy: 32.9 -> 56.4 | Improved |
| add all three proposed training techniques (ablation) | Bi-Real-34 · ImageNet val | top-1 accuracy: 53.1 -> 62.2 | Improved |
| third-order instead of second-order sign derivative approximation (ablation) | Bi-Real-18 · ImageNet val | top-1 accuracy: 56.40 -> 56.46 | No clear effect |
| third-order instead of second-order sign derivative approximation (ablation) | Bi-Real-18 · ImageNet val | top-5 accuracy: 79.50 -> 79.74 | No clear effect |
| add the proposed training techniques to plain network (ablation) | Plain-18 · ImageNet val | top-1 accuracy: 3.3 -> 12.1 | Improved |
| add extra shortcut inside bottleneck (Bi-Real vs BinaryNet binarization) | 50-layer 1-bit ResNet · ImageNet val | top-1 accuracy: 9.4 -> 62.6 | Improved |
| add extra shortcut inside bottleneck (Bi-Real vs BinaryNet binarization) | 152-layer 1-bit ResNet · ImageNet val | top-1 accuracy: 8.9 -> 64.5 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- XNOR-Net — XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks (ECCV 2016) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Bi-Real Net: Enhancing the Performance of 1-bit CNNs with Improved Representational Capability and Advanced Training Algorithm (ECCV 2018 · similarity 0.97)
- Training binary neural networks with real-to-binary convolutions (ICLR 2020 · similarity 0.92)
- BD-Net: Has Depth-Wise Convolution Ever Been Applied in Binary Neural Networks? (AAAI 2026 · similarity 0.91)
- ReActNet: Towards Precise Binary Neural Network with Generalized Activation Functions (ECCV 2020 · similarity 0.89)
- Towards Accurate Binary Convolutional Neural Network (NeurIPS 2017 · similarity 0.89)
- High-Capacity Expert Binary Networks (ICLR 2021 · similarity 0.89)
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.
