Research map / Architectures and efficiency / Quantization
PalQuant: Accelerating High-Precision Networks on Low-Precision Accelerators
Qinghao Hu, Gang Li, Qiman Wu, Jian Cheng
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 cyclic shuffle and per-stage channel shuffle (ablation) | PalQuant B=2,G=2 · ImageNet ResNet-18 | top-1 accuracy: 67.94 -> 71.13 | Improved |
| add per-block channel shuffle alone (ablation) | PalQuant B=2,G=2 · ImageNet ResNet-18 | top-1 accuracy: 67.94 -> 70.31 | Improved |
| add cyclic shuffle alone (ablation) | PalQuant B=2,G=2 · ImageNet ResNet-18 | top-1 accuracy: 67.94 -> 70.22 | Improved |
| add cyclic shuffle to per-stage channel shuffle (ablation) | PalQuant B=2,G=2 · ImageNet ResNet-18 | top-1 accuracy: 70.24 -> 71.13 | Improved |
| use per-stage instead of per-block channel shuffle with cyclic shuffle (ablation) | PalQuant B=2,G=2 · ImageNet ResNet-18 | top-1 accuracy: 70.62 -> 71.13 | Improved |
| add cyclic permutation inside the cyclic shuffle (ablation) | PalQuant B=2,G=2 · ImageNet ResNet-18 | top-1 accuracy: 70.37 -> 71.13 | Improved |
| add the whole cyclic shuffle module (ablation) | PalQuant B=2,G=2 · ImageNet ResNet-18 | top-1 accuracy: 70.23 -> 71.13 | Improved |
| add cyclic shuffle to ShuffleNet-v2 | ShuffleNet v2 · ImageNet | top-1 accuracy: 68.71 -> 69.19 | Improved |
| parallel low-precision groups (G=2) instead of LSQ 4b | PalQuant vs LSQ · ImageNet ResNet-18 | top-1 accuracy: 70.70 -> 71.12 | Improved |
| increase groups G from 2 to 4 (ablation) | PalQuant B=2 · ImageNet ResNet-18 | top-1 accuracy: 71.12 -> 72.74 | Improved |
| PalQuant instead of LSQ at matched BitOps (6b A,2b W) | PalQuant B=2,G=3 · ImageNet ResNet-18 | top-1 accuracy: 69.47 -> 72.36 | Improved |
| increase B from 2 to 4 with G=2 (ablation) | PalQuant · ImageNet ResNet-18 | top-1 accuracy: 71.13 -> 73.48 | Improved |
Closest papers by meaning
- PikeLPN: Mitigating Overlooked Inefficiencies of Low-Precision Neural Networks (CVPR 2024 · similarity 0.86)
- BatchQuant: Quantized-for-all Architecture Search with Robust Quantizer (NeurIPS 2021 · similarity 0.86)
- SpinQuant: LLM Quantization with Learned Rotations (ICLR 2025 · similarity 0.86)
- A Review of State-of-the-art Mixed-Precision Neural Network Frameworks (TPAMI 2024 · similarity 0.86)
- Non-uniform Step Size Quantization for Accurate Post-Training Quantization (ECCV 2022 · similarity 0.86)
- Towards Cheaper Inference in Deep Networks with Lower Bit-Width Accumulators (ICLR 2024 · 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.
