Research map / Architectures and efficiency / Quantization
Precision Gating: Improving Neural Network Efficiency with Dynamic Dual-Precision Activations
Yichi Zhang, Ritchie Zhao, Weizhe Hua, Nayun Xu, G. Edward Suh, Zhiru Zhang
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 dynamic dual-precision gating instead of 2-bit uniform quantization | ShiftNet-20 · CIFAR-10 | top-1 accuracy: 77.8 -> 84.5 | Improved |
| Use dynamic dual-precision gating instead of 2-bit uniform quantization | ResNet-20 · CIFAR-10 | top-1 accuracy: 84.0 -> 90.6 | Improved |
| Use dynamic dual-precision gating instead of 5-bit uniform quantization | ShuffleNet V2 0.5× · ImageNet | top-1 accuracy: 57.0 -> 58.0 | Improved |
| Use dynamic dual-precision gating instead of 8-bit uniform quantization | LSTM · PTB | perplexity per word: 109.8 -> 108.5 | Improved |
| Replace the learnable gate with a fixed threshold (ablation) | Precision Gating, ResNet-20 · CIFAR-10, 3/2 | top-1 accuracy: 91.2 -> 88.7 | Hurt |
| Add sparse back-propagation (ablation) | Precision Gating, ShiftNet-20 · CIFAR-10 | sparsity: 49.3 -> 55.5; 84.0 -> 96.3 | Improved |
| Add sparse back-propagation (ablation) | Precision Gating, ResNet-20 · CIFAR-10 | sparsity: 58.2 -> 78.2; 76.5 -> 90.1 | Improved |
| Add sparse back-propagation (ablation) | Precision Gating, ShuffleNet V2 0.5× · ImageNet | sparsity: 36.6 -> 57.2; 43.1 -> 62.2 | Improved |
Closest papers by meaning
- Not All Bits have Equal Value: Heterogeneous Precisions via Trainable Noise (NeurIPS 2022 · similarity 0.88)
- Towards Mixed-Precision Quantization of Neural Networks via Constrained Optimization (ICCV 2021 · similarity 0.87)
- DQT: Dynamic Quantization Training via Dequantization-Free Nested Integer Arithmetic (AAAI 2026 · similarity 0.87)
- LEARNED STEP SIZE QUANTIZATION (ICLR 2020 · similarity 0.87)
- Effective Training of Convolutional Neural Networks With Low-Bitwidth Weights and Activations (TPAMI 2021 · similarity 0.87)
- PikeLPN: Mitigating Overlooked Inefficiencies of Low-Precision Neural Networks (CVPR 2024 · similarity 0.86)
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.
