Research map / Architectures and efficiency / Quantization
Entropy-Driven Mixed-Precision Quantization for Deep Network Design
Zhenhong Sun, Ce Ge, Junyan Wang, Ming Lin, Hesen Chen, Hao Li, Xiuyu Sun
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 |
|---|---|---|---|
| Quantization Bits Refinement (QBR) (ablation) | Ours-mixed · 256kB SRAM, 1MB Flash | TOP-1 accuracy: 65.6% -> 66.5% | Improved |
| Quantization Bits Refinement (QBR) (ablation) | Ours-mixed · 320kB SRAM, 1MB Flash | TOP-1 accuracy: 66.6% -> 68.2% | Improved |
| Quantization Bits Refinement (QBR) (ablation) | Ours-mixed · 512kB SRAM, 2MB Flash | TOP-1 accuracy: 71.7% -> 72.8% | Improved |
| Mixed precision instead of fixed 4-bit (ablation) | Ours · 256kB SRAM, 1MB Flash | TOP-1 accuracy: 64.4% -> 66.5% | Improved |
| Mixed precision instead of fixed 8-bit (ablation) | Ours · 256kB SRAM, 1MB Flash | TOP-1 accuracy: 62.7% -> 66.5% | Improved |
| Fixed 4-bit instead of fixed 8-bit (ablation) | Ours · 512kB SRAM, 2MB Flash | TOP-1 accuracy: 70.4% -> 71.0% | Improved |
| Ours-19.2G search vs MobileNetV2 8-bit | MobileNetV2 · ImageNet, 19.2G BitOps | TOP-1 accuracy: 71.9% -> 74.8% | Improved |
| Ours-7.0G search vs MobileNetV2 4-bit | MobileNetV2 · ImageNet, 7.0G BitOps | TOP-1 accuracy: 68.9% -> 70.8% | Improved |
| Ours vs MCUNet-int8 | MCUNet · 256kB SRAM, 1MB Flash | TOP-1 accuracy: 60.3% -> 66.5% | Improved |
| Ours vs MCUNet-int4 | MCUNet · 256kB SRAM, 1MB Flash | TOP-1 accuracy: 62.0% -> 66.5% | Improved |
| Ours vs MCUNetV2 | MCUNetV2 · 512kB SRAM, 2MB Flash | TOP-1 accuracy: 71.8% -> 72.8% | Improved |
| Ours-Face vs MCUNetV2 on WIDER FACE | MCUNetV2 · WIDER FACE Hard | mAP: 0.55 -> 0.77 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- MCUNetV2 — MCUNet: Tiny Deep Learning on IoT Devices (NeurIPS 2020) · better on 3 of 5 shared results
- RNNPool — RNNPool: Efficient Non-linear Pooling for RAM Constrained Inference (NeurIPS 2020) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- On-Device Training Under 256KB Memory (NeurIPS 2022 · similarity 0.88)
- Constrained deep neural network architecture search for IoT devices accounting for hardware calibration (NeurIPS 2019 · similarity 0.88)
- Efficient and Generalizable Mixed-Precision Quantization via Topological Entropy (NeurIPS 2025 · similarity 0.87)
- MCUNet: Tiny Deep Learning on IoT Devices (NeurIPS 2020 · similarity 0.86)
- Towards Mixed-Precision Quantization of Neural Networks via Constrained Optimization (ICCV 2021 · similarity 0.86)
- TinyTrain: Resource-Aware Task-Adaptive Sparse Training of DNNs at the Data-Scarce Edge (ICML 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.
