Research map / Architectures and efficiency / Quantization
Finding the Task-Optimal Low-Bit Sub-Distribution in Deep Neural Networks
Runpei Dong, Zhanhong Tan, Mengdi Wu, Linfeng Zhang, Kaisheng Ma
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 task-objective DGMS quantization instead of LQNets | ResNet-20 and VGG-Small · CIFAR-10, 3-bit | Top-1 accuracy: 92.00% -> 92.84%; 93.80% -> 94.46% | Improved |
| Use DGMS quantization instead of TTQ | ResNet-32 and ResNet-56 · CIFAR-10, 2-bit | Top-1 accuracy: 92.37% -> 92.97%; 93.56% -> 93.72% | Improved |
| Use DGMS quantization instead of LQNets | ResNet-20 and VGG-Small · CIFAR-10, 2-bit | Top-1 accuracy: 91.80% -> 92.13%; 93.80% -> 94.36% | Improved |
| Use DGMS weight quantization instead of the listed baselines | ResNet-18 · ImageNet, 4/32 and 4/4 | Top-1 accuracy: 70.00% -> 70.25%; 68.45% -> 68.95% | Improved |
| Use DGMS weight quantization instead of HAQ | ResNet-50 · ImageNet, 2/32 | Top-1 accuracy: 70.63% -> 72.88% | Improved |
| Use DGMS weight quantization instead of HAQ | ResNet-50 · ImageNet, 4/32 | Top-1 accuracy: 76.14% -> 76.28% | Improved |
| Quantize detector weights with DGMS | SSDLite-MBV2 and SSDLite-MBV3 · PASCAL VOC2007 test | mAP: 68.6 -> 67.9; 67.2 -> 65.7 | Hurt |
| Deploy 4-bit models with Q-SIMD rather than primitive TVM scheduling | DGMS quantized models · Inference on Qualcomm 888 | speedup: 7.46×; 4.34×; 1.74×; 1.66×; 1.72× | Improved |
| Initialize gamma empirically rather than with standard deviation (ablation) | DGMS · CIFAR-10, ResNet-20 2-bit | Top-1 accuracy: 89.26% -> 92.13% | Improved |
| Initialize gamma empirically rather than with standard deviation (ablation) | DGMS · CIFAR-10, VGG-Small 2-bit | Top-1 accuracy: 92.55% -> 94.41% | Improved |
| Learn temperature instead of fixing it at initialization 0.010 (ablation) | DGMS on VGG-Small 3-bit · CIFAR-10 | Top-1 accuracy: 94.56% -> 94.64% | No clear effect |
| Increase the number of Gaussian mixture components (ablation) | DGMS on ResNet-18 4-bit · ImageNet | Top-1 accuracy | 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.
- LQNets 4/32 — LQ-Nets: Learned Quantization for Highly Accurate and Compact Deep Neural Networks (ECCV 2018) · 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
- Quantization Networks (CVPR 2019 · similarity 0.89)
- Data-Free Quantization Through Weight Equalization and Bias Correction (ICCV 2019 · similarity 0.89)
- Deep Neural Network Quantization via Layer-Wise Optimization Using Limited Training Data (AAAI 2019 · similarity 0.89)
- Explicit Loss-Error-Aware Quantization for Low-Bit Deep Neural Networks (CVPR 2018 · similarity 0.88)
- Distribution-Aware Adaptive Multi-Bit Quantization (CVPR 2021 · similarity 0.88)
- QBB: Quantization with Binary Bases for LLMs (NeurIPS 2024 · similarity 0.88)
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.
