Research map / Architectures and efficiency / Quantization
BASQ: Branch-Wise Activation-Clipping Search Quantization for Sub-4-Bit Neural Networks
Han-Byul Kim, Eunhyeok Park, Sungjoo Yoo
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 BASQ activation quantization instead of LSQ (ablation) | 2-bit ResNet-20 · CIFAR10 | top-1 accuracy (%): 89.40 -> 90.21 | Improved |
| use BASQ activation quantization instead of LSQ (ablation) | 2-bit MobileNet-v2 · CIFAR10 | top-1 accuracy (%): 89.62 -> 90.47 | Improved |
| add new building block with dedicated BatchNorm (ablation) | 2-bit ResNet-20 · CIFAR10 | top-1 accuracy (%): 89.40 -> 89.72 | Improved |
| add new building block with dedicated BatchNorm (ablation) | 2-bit MobileNet-v2 · CIFAR10 | top-1 accuracy (%): 89.62 -> 89.74 | Improved |
| use FracBNN block instead of basic block (ablation) | 2-bit ResNet-20 · CIFAR10 | top-1 accuracy (%): 89.40 -> 89.39 | No clear effect |
| use FracBNN block instead of basic block (ablation) | 2-bit MobileNet-v2 · CIFAR10 | top-1 accuracy (%): 89.62 -> 89.32 | Hurt |
| add all three Flexconn skip connections with new block (ablation) | 2-bit MobileNet-v2 · CIFAR10 | top-1 accuracy (%): 89.74 -> 90.94 | Improved |
| share convolution across branches instead of private (ablation) | BASQ 2-bit ResNet-20 · CIFAR10 | top-1 accuracy (%): 89.90 -> 90.21 | Improved |
| search L2 decay weight instead of random selection (ablation) | 2-bit MobileNet-v2 · ImageNet | top-1 accuracy (%): [0.08%, 57.76%] range of random archs vs 64.71 | Improved |
| random L2 decay selection (ablation) | 2-bit ResNet-18 · ImageNet | top-1 accuracy (%): [67.98%, 68.53%] | No clear effect |
| train LSQ with BASQ 2-phase schedule, compared to BASQ (ablation) | 2-bit MobileNet-v2 · ImageNet | top-1 accuracy (%): 50.0 -> 64.7 | Improved |
| train LSQ with BASQ 2-phase schedule, compared to BASQ (ablation) | 4-bit MobileNet-v2 · ImageNet | top-1 accuracy (%): 70.2 -> 72.0 | Improved |
Closest papers by meaning
- PROFIT: A Novel Training Method for sub-4-bit MobileNet Models (ECCV 2020 · similarity 0.88)
- MetaMix: Meta-State Precision Searcher for Mixed-Precision Activation Quantization (AAAI 2024 · similarity 0.87)
- Pushing the Limit of Post-Training Quantization (TPAMI 2025 · similarity 0.87)
- Once Quantization-Aware Training: High Performance Extremely Low-Bit Architecture Search (ICCV 2021 · similarity 0.86)
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.86)
- JAQ: Joint Efficient Architecture Design and Low-Bit Quantization with Hardware-Software Co-Exploration (AAAI 2025 · 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.
