Research map / Architectures and efficiency / Quantization
HAWQ-V3: Dyadic Neural Network Quantization
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 HAWQ-V3 integer-only INT8 quantization instead of the Integer Only baseline | ResNet50 · ImageNet | Top-1 accuracy: 74.90 -> 77.58 | Helped |
| Use HAWQ-V3 integer-only INT8 quantization instead of the Integer Only baseline | InceptionV3 · ImageNet | Top-1 accuracy: 74.20 -> 78.76 | Helped |
| Use mixed INT4/INT8 quantization instead of uniform INT4 | ResNet18 · ImageNet | Top-1 accuracy: 68.45 -> 70.22 | Helped |
| Add distillation to mixed-precision quantization (ablation) | ResNet50 · ImageNet | Top-1 accuracy: 75.39 -> 76.73 | Helped |
| Add distillation to mixed-precision quantization (ablation) | ResNet18 · ImageNet | Top-1 accuracy: 70.22 -> 70.38 | Helped |
| Add distillation to mixed-precision quantization (ablation) | InceptionV3 · ImageNet | Top-1 accuracy: 74.65 -> 74.72 | Helped |
| Add distillation to uniform INT8 or uniform INT4 quantization (ablation) | HAWQ-V3 · ImageNet models | Top-1 accuracy | No clear effect |
| Use hardware-measured latency as an ILP constraint when choosing layer bit-widths | HAWQ-V3 mixed precision · ImageNet, ResNet50 | speedup over INT8: 1x -> 1.23x | Helped |
| Replace uniform INT8 with uniform INT4 | HAWQ-V3 · ImageNet, ResNet50 | speedup over INT8: 1x -> 1.45x | Helped |
| Use the proposed BN folding procedure: update BN statistics before freezing and folding (ablation) | HAWQ-V3 · ImageNet quantization | Top-1 accuracy | Helped |
| Keep residual connections in floating point rather than using integer-only residual arithmetic | Quantized residual networks · Integer-only deployment | signal mismatch: more than 90% | Hurt |
Closest papers by meaning
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.88)
- PTMQ: Post-training Multi-Bit Quantization of Neural Networks (AAAI 2024 · similarity 0.88)
- A2Q+: Improving Accumulator-Aware Weight Quantization (ICML 2024 · similarity 0.88)
- Efficient Multi-bit Quantization Network Training via Weight Bias Correction and Bit-wise Coreset Sampling (NeurIPS 2025 · similarity 0.88)
- Towards Cheaper Inference in Deep Networks with Lower Bit-Width Accumulators (ICLR 2024 · similarity 0.87)
- JAQ: Joint Efficient Architecture Design and Low-Bit Quantization with Hardware-Software Co-Exploration (AAAI 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.