Research map / Architectures and efficiency / Quantization
Bayesian Bits: Unifying Quantization and Pruning
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 |
|---|---|---|---|
| Increase the regularization strength from μ=0.01 to μ=0.1 | Bayesian Bits · CIFAR10 test | Rel. GBOPs (%): 0.51±0.03 -> 0.29±0.00 | Helped |
| Increase the regularization strength from μ=0.01 to μ=0.1 | Bayesian Bits · CIFAR10 test | accuracy (%): 93.23±0.10 -> 91.96±0.04 | Hurt |
| Increase the regularization strength from μ=0.01 to μ=0.2 | Bayesian Bits · ImageNet validation | Rel. GBOPs (%): 2.85±0.04 -> 0.68±0.03 | Helped |
| Increase the regularization strength from μ=0.01 to μ=0.2 | Bayesian Bits · ImageNet validation | Top-1 accuracy (%): 69.84±0.02 -> 63.76±0.34 | Hurt |
| Disable pruning and learn quantization gates only (ablation) | Bayesian Bits · ImageNet validation | Rel. GBOPs (%): 1.93±0.05 -> 1.78 ± 0.11 | Hurt |
| Replace stochastic gates with deterministic gates (ablation) | Bayesian Bits · ImageNet validation, ResNet18, μ=0.03 | Top-1 accuracy (%): 69.36±0.11 -> 56.81 | Hurt |
| Replace stochastic gates with deterministic gates (ablation) | Bayesian Bits · CIFAR10 test, VGG, μ=0.01 | accuracy (%): 93.23±0.10 -> 92.82 | Hurt |
| Use stochastic gates instead of deterministic gates (ablation) | Bayesian Bits · ImageNet validation, ResNet18, μ=0.03 | Top-1 accuracy (%): 56.81 -> 69.36±0.11 | Helped |
| Fine-tune weights and quantization ranges after fixing gates (ablation) | Bayesian Bits · ImageNet validation, ResNet18, μ=0.2 | Top-1 accuracy (%): 62.32±0.71 -> 63.76±0.34 | Helped |
| Leave activations feeding into residual connections unquantized (ablation) | Bayesian Bits · ImageNet validation, ResNet18 | Rel. GBOPs (%): 2.85±0.04 -> 2.41 ± 0.03 | Helped |
Closest papers by meaning
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.88)
- ScaleBITS: Scalable Bitwidth Search for Hardware-Aligned Mixed-Precision LLMs (NeurIPS 2026 · similarity 0.87)
- Retraining-Free Model Quantization via One-Shot Weight-Coupling Learning (CVPR 2024 · similarity 0.87)
- Q-Palette: Fractional-Bit Quantizers Toward Optimal Bit Allocation for Efficient LLM Deployment (NeurIPS 2025 · similarity 0.87)
- Learning from Loss Landscape: Generalizable Mixed-Precision Quantization via Adaptive Sharpness-Aware Gradient Aligning (ICML 2025 · similarity 0.87)
- QBB: Quantization with Binary Bases for LLMs (NeurIPS 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.