Research map / Architectures and efficiency / Quantization
ProxyBNN: Learning Binarized Neural Networks via Proxy Matrices
Xiangyu He, Zitao Mo, Ke Cheng, Weixiang Xu, Qinghao Hu, Peisong Wang, Qingshan Liu, Jian 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 |
|---|---|---|---|
| use orthogonal proxy matrix instead of random matrix (ablation) | ProxyBNN ResNet18 · CIFAR-100 (from scratch) | top-1 accuracy: 53.58 -> 67.17 | Improved |
| use orthogonal proxy matrix instead of MSE matrix (ablation) | ProxyBNN ResNet18 · CIFAR-100 (from scratch) | top-1 accuracy: 30.69 -> 67.17 | Improved |
| use orthogonal proxy matrix instead of random matrix (ablation) | ProxyBNN ResNet-18 · ImageNet | top-1 accuracy: 53.7 -> 58.7 | Improved |
| use orthogonal proxy matrix instead of MSE matrix (ablation) | ProxyBNN ResNet-18 · ImageNet | top-1 accuracy: 36.8 -> 58.7 | Improved |
| use orthogonal proxy matrix instead of random matrix (ablation) | ProxyBNN WRN22 · CIFAR-10 (from scratch) | top-1 accuracy: 91.24 -> 92.96 | Improved |
| minimize quantization error only (MSE basis) despite 8x smaller average quantization error than Bi-Real (ablation) | ProxyBNN WRN22 · CIFAR-100 | top-1 accuracy: 69.73 -> 59.32 | Hurt |
| set penalty factor gamma from 0.001 to 1e-5 (ablation) | ProxyBNN WRN22 · CIFAR-100 | error rate: 30.28 -> 28.43 | Improved |
| decrease penalty factor gamma from 1e-5 to 1e-7 (ablation) | ProxyBNN WRN22 · CIFAR-100 | error rate: 28.43 -> 31.33 | Hurt |
| set proxy matrix learning rate to lr_w x 0.1 instead of lr_w (ablation) | ProxyBNN WRN22 · CIFAR-100 | error rate: 30.77 -> 28.43 | Improved |
| reduce proxy matrix learning rate further to lr_w x 0.001 (ablation) | ProxyBNN WRN22 · CIFAR-100 | error rate: 28.43 -> 31.44 | Hurt |
| use ProxyBNN instead of SiBNN with modified ResNet-18 (1-bit/1-bit) | ResNet-18 (modified) · ImageNet | top-1 accuracy: 59.7 -> 63.3 | Improved |
| use ProxyBNN instead of XNOR++ with ResNet-18 (1-bit/1-bit) | ResNet-18 · ImageNet | top-1 accuracy: 57.1 -> 58.7 | Improved |
Later papers that compared with this method
1 later accepted paper reports ProxyBNN as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Lipschitz Continuity Retained Binary Neural Network (ECCV 2022 · 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
- Training Binary Neural Networks via Gaussian Variational Inference and Low-Rank Semidefinite Programming (NeurIPS 2024 · similarity 0.89)
- Understanding Neural Network Binarization with Forward and Backward Proximal Quantizers (NeurIPS 2023 · similarity 0.88)
- Fast and Slow Gradient Approximation for Binary Neural Network Optimization (AAAI 2025 · similarity 0.86)
- HADAMRNN: BINARY AND SPARSE TERNARY ORTHOGONAL RNNS (ICLR 2025 · similarity 0.86)
- BiPer: Binary Neural Networks using a Periodic Function (CVPR 2024 · similarity 0.85)
- Taming Binarized Neural Networks and Mixed-Integer Programs (AAAI 2024 · 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.
