Research map / Architectures and efficiency / Quantization
Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 integer-only 8-bit quantization with quantization-aware training | ResNet-50 · ImageNet | accuracy: 76.4% -> 74.9% | Hurt |
| Use the proposed 8-bit weight-and-activation quantization instead of FGQ | ResNet-50 · ImageNet | accuracy: 70.8% -> 74.9% | Helped |
| Reduce quantization from 8 bits to 7 bits (ablation) | Inception v3 with ReLU6 · ImageNet | accuracy: 75.4% -> 75.0% | Hurt |
| Use ReLU6 rather than ReLU (ablation) | Quantized Inception v3 · ImageNet | accuracy: 74.2% -> 75.4% (8 bits); 73.7% -> 75.0% (7 bits) | Helped |
| Quantize MobileNet SSD to 8 bits rather than use floating-point inference | MobileNet SSD, depth multiplier 100% · COCO | latency on Snapdragon 835 LITTLE: 778 -> 687 | Helped |
| Quantize MobileNet SSD to 8 bits rather than use floating-point inference | MobileNet SSD, depth multiplier 100% · COCO | mAP: 22.1 -> 21.7 | Hurt |
| Quantize the face detector to 8 bits rather than use floating-point inference | MobileNet SSD, depth multiplier 100% · Face detection | latency on Snapdragon 835 LITTLE, 1 core: 711 -> 372 | Helped |
| Quantize the face detector to 8 bits rather than use floating-point inference | MobileNet SSD, depth multiplier 100% · Face detection | precision: 68% -> 66% | Hurt |
| Reduce both weights and activations from 8 bits to 4 bits (ablation) | MobileNet face-attribute classifier · Face attributes | relative average category precision: -0.9% -> -14.0% | Hurt |
| Reduce both weights and activations from 8 bits to 4 bits (ablation) | MobileNet face-attribute classifier · Face attributes | age precision: -1.3% -> -19.5% | Hurt |
| Reduce weight precision while keeping activations at 8 bits (from 8/8 to 4/8) (ablation) | MobileNet face-attribute classifier · Face attributes | relative average category precision: -0.9% -> -11.4% | Hurt |
| Reduce activation precision while keeping weights at 8 bits (from 8/8 to 8/4) (ablation) | MobileNet face-attribute classifier · Face attributes | relative average category precision: -0.9% -> -3.5% | Hurt |
Closest papers by meaning
- Linear Symmetric Quantization of Neural Networks for Low-precision Integer Hardware (ICLR 2020 · similarity 0.89)
- DQT: Dynamic Quantization Training via Dequantization-Free Nested Integer Arithmetic (AAAI 2026 · similarity 0.89)
- Towards Efficient Verification of Quantized Neural Networks (AAAI 2024 · similarity 0.87)
- A2Q+: Improving Accumulator-Aware Weight Quantization (ICML 2024 · similarity 0.85)
- Scaling Laws for Floating–Point Quantization Training (ICML 2025 · similarity 0.85)
- GradQ-ViT: Robust and Efficient Gradient Quantization for Vision Transformers (AAAI 2025 · 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.