Research map / Architectures and efficiency / Quantization
Genie: Show Me the Data for 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 Genie-D synthetic data and Genie-M quantization instead of the ZeroQ+QDrop baseline (ablation) | CNN quantization · ImageNet, W4A4 and W2A4 | top-1 accuracy (%): 69.19 -> 69.66; 61.96 -> 65.10 | Helped |
| Add Genie-M to the full Genie-D configuration (ablation) | Genie-D · ImageNet (ResNet-18), W4A4 and W2A4 | top-1 accuracy (%): 69.62 -> 69.66; 64.34 -> 65.10 | Helped |
| Add trained latent vectors to the generator (ablation) | Genie-D · ImageNet (ResNet-18), W4A4 and W2A4 | top-1 accuracy (%): 69.17 -> 69.58; 60.13 -> 64.06 | Helped |
| Add swing convolution to generator-plus-latent data distillation (ablation) | Genie-D · ImageNet (ResNet-18), W4A4 and W2A4 | top-1 accuracy (%): 69.58 -> 69.62; 64.06 -> 64.34 | Helped |
| Use swing convolution for direct image distillation instead of the baseline distiller (ablation) | Zero-shot quantization · ImageNet (ResNet-18), W4A4 and W2A4 | top-1 accuracy (%): 69.19 -> 69.49; 61.96 -> 63.74 | Helped |
| Add Genie-M to the ZeroQ+QDrop baseline (ablation) | Zero-shot quantization · ImageNet (ResNet-18), W4A4 and W2A4 | top-1 accuracy (%): 69.19 -> 69.25; 61.96 -> 62.62 | Helped |
| Replace QDrop with Genie-M while retaining Genie-D data (ablation) | Genie-D · ImageNet (ResNet-18, ResNet-50, MobileNetV2), W4A4 | top-1 accuracy (%): 69.62 -> 69.66; 75.45 -> 75.59; 68.84 -> 68.38 | Mixed |
| Use Genie instead of QDrop | CNN quantization · ImageNet (ResNet-18, ResNet-50, MobileNetV2), W2A4 | top-1 accuracy (%): 65.25 -> 65.10; 70.65 -> 69.99; 54.22 -> 53.38 | Mixed |
| Use Genie instead of QDrop | CNN quantization · ImageNet (ResNet-18, ResNet-50, MobileNetV2), W4A4 | top-1 accuracy (%): 69.62 -> 69.66; 75.45 -> 75.59; 68.84 -> 68.38 | Mixed |
| Use Genie-M with QDrop rather than Genie-M without QDrop (ablation) | Genie-M · ImageNet sampled real data, W2A4 | top-1 accuracy (%): 65.27 -> 65.77; 70.39 -> 70.51; 55.55 -> 56.38; 63.66 -> 64.55; 71.79 -> 72.35; 62.76 -> 64.10 | Helped |
| Use Genie-M with QDrop rather than Genie-M without QDrop (ablation) | Genie-M · ImageNet sampled real data, W4A4 | top-1 accuracy (%): 69.13 -> 69.35; 74.93 -> 75.21; 68.22 -> 68.65; 70.87 -> 71.13; 76.50 -> 76.75; 72.68 -> 73.37 | Helped |
| Use PTQ with Genie instead of QAT with Genie-D+AIT | ResNet-18 · ImageNet, W4A4 | top-1 accuracy (%): 63.98 -> 68.51 | Helped |
Closest papers by meaning
- Sharpness-Aware Data Generation for Zero-shot Quantization (ICML 2024 · similarity 0.91)
- SynQ: Accurate Zero-shot Quantization by Synthesis-aware Fine-tuning (ICLR 2025 · similarity 0.88)
- Zero-Shot Quantization via Weight-Space Arithmetic (NeurIPS 2026 · similarity 0.87)
- Quantization without Tears (CVPR 2025 · similarity 0.86)
- One-Step Forward and Backtrack: Overcoming Zig-Zagging in Loss-Aware Quantization Training (AAAI 2024 · similarity 0.86)
- Enhancing Diversity for Data-free Quantization (CVPR 2025 · 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.