Research map / Architectures and efficiency / Quantization
Differentiable Dynamic Quantization with Mixed Precision and Adaptive Resolution
Zhaoyang Zhang, Wenqi Shao, Jinwei Gu, Xiaogang Wang, Ping Luo
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 DDQ for 4-bit weights and 8-bit activations instead of uniform quantization | MobileNetV2 · ImageNet | Top-1 accuracy: 67.1 -> 71.2 | Improved |
| Use DDQ for 4-bit weights and 8-bit activations instead of power-of-two quantization | MobileNetV2 · ImageNet | Top-1 accuracy: 69.2 -> 71.2 | Improved |
| Add gradient correction to fixed-precision DDQ (ablation) | MobileNetV2 · ImageNet, 4/8-bit | Top-1 accuracy: 71.2 -> 71.6 | Improved |
| Add gradient correction to fixed-precision DDQ (ablation) | ResNet18 · ImageNet, 4/8-bit | Top-1 accuracy: 70.8 -> 70.9 | Improved |
| Use mixed rather than fixed bitwidth under the 4-bit target (ablation) | DDQ on MobileNetV2 · ImageNet | Top-1 accuracy: 70.7 -> 71.5 | Improved |
| Use mixed rather than fixed bitwidth under the 4-bit target (ablation) | DDQ on ResNet18 · ImageNet | Top-1 accuracy: 70.7 -> 71.0 | Improved |
| Learn adaptive quantization levels instead of using uniform levels (ablation) | DDQ on MobileNetV2 · ImageNet, 4/4-bit | Top-1 accuracy: 65.2 -> 70.7 | Improved |
| Learn adaptive quantization levels instead of using power-of-two levels (ablation) | DDQ on MobileNetV2 · ImageNet, 4/4-bit | Top-1 accuracy: 68.6 -> 70.7 | Improved |
| Use DDQ with 2/2-bit quantization (ablation) | MobileNetV2 · ImageNet | Top-1 accuracy: 71.9 -> 55.4 | Hurt |
| Use DDQ mixed precision on mobile DSP instead of full precision | MobileNetV2 · Mobile DSP | Latency: 7.8 -> 5.1 | Improved |
| Use DDQ mixed precision instead of uniform quantization | MobileNetV2 · Mobile DSP | Top-1 accuracy: 67.1 -> 71.9 | Improved |
| Use mixed precision with uniform quantizer constraints instead of original DDQ (ablation) | MobileNetV2 · Mobile DSP | Top-1 accuracy: 71.9 -> 70.8 | Hurt |
Closest papers by meaning
- SDQ: Stochastic Differentiable Quantization with Mixed Precision (ICML 2022 · similarity 0.90)
- Differentiable Soft Quantization: Bridging Full-Precision and Low-Bit Neural Networks (ICCV 2019 · similarity 0.89)
- DBQ: A Differentiable Branch Quantizer for Lightweight Deep Neural Networks (ECCV 2020 · similarity 0.87)
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.85)
- Learning from Loss Landscape: Generalizable Mixed-Precision Quantization via Adaptive Sharpness-Aware Gradient Aligning (ICML 2025 · similarity 0.84)
- QBB: Quantization with Binary Bases for LLMs (NeurIPS 2024 · similarity 0.84)
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.
