Research map / Architectures and efficiency / Quantization
Sub-Bit Neural Networks: Learning To Compress and Accelerate Binary Neural Networks
Yikai Wang, Yi Yang, Fuchun Sun, Anbang Yao
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 subsets refinement (ablation) | Vanilla-SNN 0.56-bit · ImageNet ResNet-18 (W/A 1/32) | top-1 accuracy: 62.8 -> 63.4 | Improved |
| add subsets refinement (ablation) | Vanilla-SNN 0.44-bit · ImageNet ResNet-34 (W/A 1/1) | top-1 accuracy: 58.1 -> 58.6 | Improved |
| add subsets refinement (ablation) | Vanilla-SNN 0.44-bit · CIFAR10 ResNet-20 (W/A 0.44/1) | top-1 accuracy: 82.0 -> 82.5 | Improved |
| use layer-specific subsets instead of layer-shared (ablation) | SNN 0.11-bit · CIFAR10 ResNet-20 (W/A 0.11/32) | top-1 accuracy: 73.5 -> 81.6 | Improved |
| use layer-specific subsets instead of layer-shared (ablation) | Vanilla-SNN 0.56-bit · CIFAR10 ResNet-20 (W/A 0.56/32) | top-1 accuracy: 82.6 -> 87.8 | Improved |
| use layer-specific subsets instead of layer-shared (ablation) | SNN 0.78-bit · CIFAR10 ResNet-20 (W/A 0.78/32) | top-1 accuracy: 90.0 -> 90.2 | Improved |
| reduce bit-width from 0.67 to 0.44 | SNN · ImageNet ResNet-18 (W/A 1/32) | top-1 accuracy: 64.7 -> 60.9 | Hurt |
| replace 1-bit IR-Net with 0.56-bit SNN | IR-Net · ImageNet ResNet-18 (W/A 1/32) | top-1 accuracy: 66.5 -> 63.4 | Hurt |
| replace 1-bit BNN with 0.56-bit SNN | 1-bit BNN · FPGA ResNet-18 | running time: 3.626 ms -> 1.159 ms (3.13x) | Improved |
| replace 1-bit BNN with 0.56-bit SNN | 1-bit BNN · FPGA ResNet-34 | running time: 7.753 ms -> 2.329 ms (3.33x) | Improved |
| select kernels uniformly by index (Baseline A) instead of Vanilla-SNN (ablation) | Vanilla-SNN · ImageNet ResNet-18 (W/A 0.56/32) | top-1 accuracy: 62.8 -> 60.5 | Hurt |
| select top frequent kernels each iteration (Baseline B) instead of Vanilla-SNN (ablation) | Vanilla-SNN · ImageNet ResNet-18 (W/A 0.56/32) | top-1 accuracy: 62.8 -> 62.7 | No clear effect |
Closest papers by meaning
- BD-Net: Has Depth-Wise Convolution Ever Been Applied in Binary Neural Networks? (AAAI 2026 · similarity 0.90)
- S$^2$NN: Sub-bit Spiking Neural Networks (NeurIPS 2025 · similarity 0.89)
- Training binary neural networks with real-to-binary convolutions (ICLR 2020 · similarity 0.89)
- Compacting Binary Neural Networks by Sparse Kernel Selection (CVPR 2023 · similarity 0.88)
- Enhancing Optimization Robustness in 1-bit Neural Networks through Stochastic Sign Descent (ECCV 2024 · similarity 0.88)
- Can we get the best of both Binary Neural Networks and Spiking Neural Networks for Efficient Computer Vision? (ICLR 2024 · similarity 0.88)
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.
