Research map / Architectures and efficiency / Quantization
Vertical Layering of Quantized Neural Networks for Heterogeneous Inference
Wu Hai, Ruifei He, Haoru Tan, Xiaojuan Qi, Kaibin Huang
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 individual BN layers per bit-width instead of shared BN (ablation) | vertical-layered model · ImageNet ResNet-18 (W/A 2/2) | Top-1 accuracy (%): 33.5 -> 66.7 | Improved |
| use individual BN layers instead of individual statistics only (ablation) | vertical-layered model · ImageNet ResNet-18 (W/A 3/3) | Top-1 accuracy (%): 68.8 -> 69.9 | Improved |
| add positive compensation in downsampling (ablation) | vertical-layered model · ImageNet ResNet-18 (W/A 2/2) | Top-1 accuracy (%): 61.3 -> 66.7 | Improved |
| add positive compensation in downsampling (ablation) | vertical-layered model · ImageNet MobileNetV2 (W/A 2/2) | Top-1 accuracy (%): 40.4 -> 50.6 | Improved |
| use unitary scalarization instead of multi-gradient descent (ablation) | vertical-layered model · ImageNet ResNet-18 (W/A 4/4) | Top-1 accuracy (%): 69.8 -> 70.6 | Improved |
| add cosine-similarity self-KD (ablation) | vertical-layered model · ImageNet ResNet-50 (W/A 2/2) | Top-1 accuracy (%): 72.4 -> 73.2 | Improved |
| add KL-divergence self-KD instead of cosine similarity (ablation) | vertical-layered model · ImageNet ResNet-18 (W/A 4/4) | Top-1 accuracy (%): 70.6 -> 70.4 | No clear effect |
| replace full-precision first and last layers with bit-width specific ones (ablation) | vertical-layered model · ImageNet ResNet-50 (W/A 2/2) | Top-1 accuracy (%): 72.4 -> 71.0 | Hurt |
| replace full-precision first and last layers with bit-width specific ones (ablation) | vertical-layered model · ImageNet ResNet-18 (W/A 2/2) | Top-1 accuracy (%): 66.7 -> 66.9 | No clear effect |
| use joint QAT instead of ascending layer training | vertical-layered model · ImageNet ResNet-18 (W/A 3/3) | Top-1 accuracy (%): 67.9 -> 69.9 | Improved |
| use joint training instead of QAT + fine-tuning | scalable bit-width model · ImageNet ResNet-50 (W/A 2/2) | Top-1 accuracy (%): 60.1 -> 72.4 | Improved |
| use vertical-layered joint model instead of bit-width specific LSQ | quantized ResNet-18 · CIFAR-100 (W/A 2/2) | Top-1 accuracy (%): 76.9 -> 76.3 | Hurt |
Closest papers by meaning
- Layer-wise Quantization for Quantized Optimistic Dual Averaging (ICML 2025 · similarity 0.90)
- Deep Neural Network Quantization via Layer-Wise Optimization Using Limited Training Data (AAAI 2019 · similarity 0.89)
- Variational Network Quantization (ICLR 2018 · similarity 0.89)
- Training Quantized Nets: A Deeper Understanding (NeurIPS 2017 · similarity 0.89)
- Once Quantization-Aware Training: High Performance Extremely Low-Bit Architecture Search (ICCV 2021 · similarity 0.88)
- Efficient Multi-bit Quantization Network Training via Weight Bias Correction and Bit-wise Coreset Sampling (NeurIPS 2025 · 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.
