Research map / Architectures and efficiency / Neural architecture search
BatchQuant: Quantized-for-all Architecture Search with Robust Quantizer
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 |
|---|---|---|---|
| two-stage elastic quantization (mix in 32-bit first) instead of single-stage (ablation) | QFA supernet · ImageNet W/A 2/2 | top-1 accuracy: 66.3% -> 68.4% | Helped |
| two-stage elastic quantization instead of single-stage (ablation) | QFA supernet · ImageNet W/A 4/4 | top-1 accuracy: 74.8% -> 75.6% | Helped |
| two-stage elastic quantization instead of single-stage (ablation) | QFA supernet · ImageNet W/A 3/3 | top-1 accuracy: 73.7% -> 74.3% | Helped |
| replace LSQ+ activation quantizer with BatchQuant (ablation) | mixed-precision supernet · ImageNet supernet training | training stability | Helped |
| remove learnable residuals gamma and beta from BatchQuant (ablation) | BatchQuant supernet · ImageNet supernet training | training stability | Helped |
| use batch-statistics mu +/- 3 sigma extreme value estimator instead of min/max or per-channel average min/max (ablation) | BatchQuant · ImageNet supernet training | training stability | Helped |
| use EMA scale estimator for shared activation quantizer (ablation) | mixed-precision supernet · MobileNet search space | training stability | Hurt |
| use fixed learnable scale per quantizer (ablation) | mixed-precision supernet · MobileNet search space | training stability | Hurt |
| joint mixed-precision supernet instead of per-bitwidth fixed-precision supernets (OQA) | QFA · ImageNet training cost | supernet training epochs: 495 -> 190 | Helped |
| QFA mixed-precision search compared to OQA at >20 MFLOPs | QFA · ImageNet top-1 vs FLOPs | top-1 accuracy | No clear effect |
| QFA mixed-precision search compared to OQA at <20 MFLOPs | QFA · ImageNet top-1 vs FLOPs | top-1 accuracy | Helped |
| NSGA-II search with no retraining instead of per-scenario search and finetune | QFA · design cost | marginal GPU hours: 24N -> 0.N | Helped |
Closest papers by meaning
- HAQ: Hardware-Aware Automated Quantization With Mixed Precision (CVPR 2019 · similarity 0.89)
- JAQ: Joint Efficient Architecture Design and Low-Bit Quantization with Hardware-Software Co-Exploration (AAAI 2025 · similarity 0.89)
- LEARNED STEP SIZE QUANTIZATION (ICLR 2020 · similarity 0.88)
- One-Shot Model for Mixed-Precision Quantization (CVPR 2023 · similarity 0.88)
- DQT: Dynamic Quantization Training via Dequantization-Free Nested Integer Arithmetic (AAAI 2026 · similarity 0.88)
- MetaMix: Meta-State Precision Searcher for Mixed-Precision Activation Quantization (AAAI 2024 · 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.