Research map / Architectures and efficiency / Quantization
Once Quantization-Aware Training: High Performance Extremely Low-Bit Architecture Search
Mingzhu Shen, Feng Liang, Ruihao Gong, Yuhang Li, Chuming Li, Chen Lin, Fengwei Yu, Junjie Yan, Wanli Ouyang
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 |
|---|---|---|---|
| search architecture on the 2-bit supernet instead of FP supernet (joint training) (ablation) | OQAT 2-bit · ImageNet 2-bit | top-1 accuracy (%): 54.4% -> 56.1% | Improved |
| search architecture on the 2-bit supernet with QAT@150 finetuning (ablation) | OQAT 2-bit · ImageNet 2-bit | top-1 accuracy (%): 28.1% -> 33.2% | Improved |
| search architecture on the 2-bit supernet with QAT@500 finetuning (ablation) | OQAT 2-bit · ImageNet 2-bit | top-1 accuracy (%): 43.5% -> 47.2% | Improved |
| initialize 2-bit supernet with bit inheritance instead of plain QAT (ablation) | OQAT 2/2 · ImageNet 2-bit | top-1 accuracy (%): 56% -> 64.5% | Improved |
| initialize 3-bit supernet with bit inheritance instead of plain QAT (ablation) | OQAT 3/3 · ImageNet 3-bit | top-1 accuracy (%): 72.1% -> 72.7% | Improved |
| bit inheritance start accuracy vs QAT start accuracy at 2-bit (ablation) | OQAT 2/2 · ImageNet 2-bit | top-1 accuracy at first epoch (%): 0.8% -> 49.3% | Improved |
| use supernet trained at 4-bit with shared step size (OQAT-4bit) instead of individual 4-bit QAT, calibrated to 3/3 (ablation) | biggest model · ImageNet 3/3 | top-1 accuracy (%): 65.3% -> 69.1% | Improved |
| calibrate OQAT-4bit supernet to 3/4 bits (ablation) | biggest model · ImageNet 3/4 | top-1 accuracy (%): 69.2% -> 72.1% | Improved |
| fine-tune the inherited 3-bit supernet for one epoch (ablation) | bit inheritance 3-bit · ImageNet 3-bit | top-1 accuracy: 2.6% | Improved |
| choose wider models at fixed floating-point accuracy | architecture analysis · sampled subnets 2-bit | quantization friendliness | Improved |
| increase input resolution at similar FLOPs | architecture analysis · sampled subnets 2-bit | quantization friendliness | Improved |
| increase network depth at fixed FLOPs | architecture analysis · sampled subnets 2-bit | quantization friendliness | Hurt |
Closest papers by meaning
- BatchQuant: Quantized-for-all Architecture Search with Robust Quantizer (NeurIPS 2021 · similarity 0.91)
- A Statistical Framework for Low-bitwidth Training of Deep Neural Networks (NeurIPS 2020 · similarity 0.91)
- Optimization-Based Post-Training Quantization With Bit-Split and Stitching (TPAMI 2022 · similarity 0.91)
- Bitwidth-Adaptive Quantization-Aware Neural Network Training: A Meta-Learning Approach (ECCV 2022 · similarity 0.90)
- JAQ: Joint Efficient Architecture Design and Low-Bit Quantization with Hardware-Software Co-Exploration (AAAI 2025 · similarity 0.90)
- Efficient Multi-bit Quantization Network Training via Weight Bias Correction and Bit-wise Coreset Sampling (NeurIPS 2025 · similarity 0.89)
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.
