Research map / Architectures and efficiency / Quantization
Differentiable Soft Quantization: Bridging Full-Precision and Low-Bit Neural Networks
Ruihao Gong, Xianglong Liu, Shenghu Jiang, Tianxiang Li, Peng Hu, Jiazhen Lin, Fengwei Yu, Junjie Yan
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 DSQ with fixed alpha (ablation) | 2-bit uniform quantization · CIFAR-10 ResNet-20, 2-bit uniform | top-1 accuracy (%): 86.63 -> 86.95 | Improved |
| learn alpha during training (evolution training) (ablation) | DSQ 2-bit · CIFAR-10 ResNet-20, 2-bit uniform | top-1 accuracy (%): 86.95 -> 87.25 | Improved |
| learn clipping values l and u jointly with alpha (ablation) | DSQ 2-bit · CIFAR-10 ResNet-20, 2-bit uniform | top-1 accuracy (%): 87.25 -> 88.44 | Improved |
| add DSQ to PACT (ablation) | PACT 2-bit · CIFAR-10 ResNet-20 | top-1 accuracy (%): 88.24 -> 90.11 | Improved |
| replace sign binarization with binary DSQ (alpha 0.2) (ablation) | binarization 1/1 · CIFAR-10 ResNet-20 | accuracy (%): 82.46 -> 83.80 | Improved |
| use piecewise DSQ instead of binary DSQ (ablation) | binarization 1/1 · CIFAR-10 ResNet-20 | accuracy (%): 83.80 -> 84.11 | Improved |
| use binary DSQ instead of plain binary (ablation) | binarization 1/32 · CIFAR-10 ResNet-20 | accuracy (%): 90.11 -> 90.24 | Improved |
| use piecewise DSQ instead of binary DSQ (ablation) | binarization 1/32 · CIFAR-10 ResNet-20 | accuracy (%): 90.24 -> 90.03 | Hurt |
| use DSQ instead of DoReFa-Net | 1-bit W/A 1/1 · CIFAR-10 ResNet-20 | accuracy (%): 79.30 -> 84.11 | Improved |
| use DSQ instead of PACT at 4-bit | MobileNetV2 · ImageNet val | accuracy (%): 61.40 -> 64.80 | Improved |
| use DSQ instead of LQ-Net at 3-bit | ResNet-34 · ImageNet val | accuracy (%): 71.90 -> 72.54 | Improved |
| use DSQ 2-bit kernels instead of NCNN 8-bit | ResNet-18 inference · Raspberry Pi 3B | time (ms): 935.51 -> 551.22 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- XNOR-Net — XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks (ECCV 2016) · better on 1 of 1 shared result
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
- SDQ: Stochastic Differentiable Quantization with Mixed Precision (ICML 2022 · similarity 0.90)
- BSQ: Exploring Bit-Level Sparsity for Mixed-Precision Neural Network Quantization (ICLR 2021 · similarity 0.90)
- Differentiable Dynamic Quantization with Mixed Precision and Adaptive Resolution (ICML 2021 · similarity 0.89)
- DBQ: A Differentiable Branch Quantizer for Lightweight Deep Neural Networks (ECCV 2020 · similarity 0.89)
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.89)
- PTMQ: Post-training Multi-Bit Quantization of Neural Networks (AAAI 2024 · similarity 0.87)
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.
