Research map / Architectures and efficiency / Quantization
BSQ: Exploring Bit-Level Sparsity for Mixed-Precision Neural Network Quantization
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 |
|---|---|---|---|
| Use BSQ mixed-precision quantization instead of HAWQ | ResNet-20 · CIFAR-10, 4-bit activation | accuracy: 92.22 -> 92.32 | Helped |
| Use BSQ mixed-precision quantization instead of HAWQ | ResNet-20 · CIFAR-10, 4-bit activation | compression rate: 13.11 -> 14.24 | Helped |
| Use BSQ instead of HAWQ | Inception-V3 · ImageNet | Top-1 accuracy: 75.52 -> 75.90 | Helped |
| Use BSQ instead of HAWQ | Inception-V3 · ImageNet | compression rate: 12.04 -> 12.89 | Helped |
| Increase regularization strength from α=3e-3 to α=2e-2 (ablation) | BSQ · CIFAR-10 (ResNet-20, 4-bit activation) | accuracy after finetuning: 92.60 -> 89.49 | Hurt |
| Increase regularization strength from α=3e-3 to α=2e-2 (ablation) | BSQ · CIFAR-10 (ResNet-20, 4-bit activation) | compression rate: 10.60 -> 36.63 | Helped |
| Add memory-consumption-aware layer-wise reweighting to the bit-level group Lasso (ablation) | BSQ · CIFAR-10 (ResNet-20, 4-bit activation) | accuracy at comparable compression | Helped |
| Remove memory-consumption-aware layer-wise reweighting (ablation) | BSQ · CIFAR-10 (ResNet-20, 4-bit activation) | compression rate | Hurt |
| Re-quantize periodically every 50 or 100 epochs rather than not during training (ablation) | BSQ · CIFAR-10 (ResNet-20) | accuracy-compression tradeoff | Helped |
| Re-quantize every 20 epochs rather than every 50 or 100 epochs (ablation) | BSQ · CIFAR-10 (ResNet-20) | training stability | Hurt |
| Increase re-quantization interval from 50 to 100 epochs (ablation) | BSQ · CIFAR-10 (ResNet-20) | accuracy across compression rates | Helped |
| Use BSQ finetuning rather than training from scratch with the same quantization scheme (ablation) | ResNet-20 · CIFAR-10, 4-bit activation | accuracy: 91.72 -> 92.60 | Helped |
Closest papers by meaning
- PTMQ: Post-training Multi-Bit Quantization of Neural Networks (AAAI 2024 · similarity 0.89)
- Retraining-Free Model Quantization via One-Shot Weight-Coupling Learning (CVPR 2024 · similarity 0.89)
- Learning from Loss Landscape: Generalizable Mixed-Precision Quantization via Adaptive Sharpness-Aware Gradient Aligning (ICML 2025 · similarity 0.88)
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.88)
- QBB: Quantization with Binary Bases for LLMs (NeurIPS 2024 · similarity 0.87)
- MetaMix: Meta-State Precision Searcher for Mixed-Precision Activation Quantization (AAAI 2024 · similarity 0.87)
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.