Research map / Architectures and efficiency / Quantization
Hardware-Centric AutoML for Mixed-Precision Quantization
Kuan Wang, Zhijian Liu, Yujun Lin, Ji Lin, Song Han
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 learned flexible per-layer bitwidths instead of uniform 4-bit PACT at same latency | MobileNet-V1 · BISMO edge latency-constrained | top-1 accuracy (%): 62.44 -> 67.40 | Improved |
| use learned flexible bitwidths instead of uniform 4-bit PACT at same latency | MobileNet-V1 · BitFusion latency-constrained | top-1 accuracy (%): 62.44 -> 67.45 | Improved |
| use learned flexible bitwidths instead of uniform 4-bit PACT at same energy | MobileNet-V1 · BitFusion energy-constrained | top-1 accuracy (%): 62.44 -> 64.78 | Improved |
| use learned flexible bitwidths instead of Deep Compression 2-bit at same model size | MobileNet-V1 · ImageNet model size constraint | top-1 accuracy (%): 37.62 -> 57.14 | Improved |
| use flexible bitwidths instead of uniform 8-bit quantization | MobileNet-V1 · BISMO edge latency | latency: 96.20 ms -> 57.70 ms | Improved |
| use flexible bitwidths instead of uniform 8-bit quantization | MobileNet-V1 · BitFusion latency | latency: 20.08 -> 11.09 ms | Improved |
| use RL instead of evolutionary search for bitwidth policy (ablation) | HAQ · BISMO edge 45 ms, MobileNet-V1 | top-1 accuracy (%): 65.73 -> 67.40 | Improved |
| use RL instead of Bayesian optimization for bitwidth policy (ablation) | HAQ · BISMO edge 45 ms, MobileNet-V1 | top-1 accuracy (%): 66.28 -> 67.40; search time 74 hours -> 17 hours | Improved |
| transfer policy agent from MobileNet-V1 to V2 instead of searching directly on V2 | HAQ · BISMO edge 4-bit latency | top-1 accuracy (%): 66.99 -> 65.80 | Hurt |
| transfer V1-trained agent to V2 versus uniform 4-bit PACT | HAQ · BISMO edge 4-bit latency | top-1 accuracy (%): 61.39 -> 65.80 | Improved |
| add direct hardware latency feedback instead of BitOPs proxy constraint (ablation) | HAQ · BISMO edge, MobileNet-V2 | latency: 85.06 ms -> 66.92 ms | Improved |
| add hardware feedback instead of BitOPs constraint (tight budget) (ablation) | HAQ · BISMO edge, MobileNet-V2 | top-1 accuracy (%): 70.29 -> 70.90 | Mixed |
Closest papers by meaning
- HAQ: Hardware-Aware Automated Quantization With Mixed Precision (CVPR 2019 · similarity 0.96)
- HAWQ-V3: Dyadic Neural Network Quantization (ICML 2021 · similarity 0.90)
- JAQ: Joint Efficient Architecture Design and Low-Bit Quantization with Hardware-Software Co-Exploration (AAAI 2025 · similarity 0.90)
- Linear Symmetric Quantization of Neural Networks for Low-precision Integer Hardware (ICLR 2020 · similarity 0.89)
- HMQ: Hardware Friendly Mixed Precision Quantization Block for CNNs (ECCV 2020 · similarity 0.89)
- ScaleBITS: Scalable Bitwidth Search for Hardware-Aligned Mixed-Precision LLMs (NeurIPS 2026 · 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.
