Research map / Architectures and efficiency / Quantization
Bitwidth-Adaptive Quantization-Aware Neural Network Training: A Meta-Learning Approach
Jiseok Youn, Jaehun Song, Hyung-Sin Kim, Saewoong Bahk
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 MEBQAT bitwidth-task meta-training instead of AdaBits | MobileNet-v2 · CIFAR-10, bitwidth (2,2) | accuracy: 58.98 -> 78.50 | Improved |
| use MEBQAT instead of dedicated per-bitwidth QAT | MobileNet-v2 · CIFAR-10, bitwidth (2,2) | accuracy: 84.40 -> 78.50 | Hurt |
| use MEBQAT instead of Any-precision DNN | 8-layer CNN · SVHN, bitwidth (2,2) | accuracy: 94.94 -> 97.25 | Improved |
| use MEBQAT instead of Any-precision DNN | 8-layer CNN · SVHN, bitwidth (4,4) | accuracy: 96.19 -> 97.62 | Improved |
| use MEBQAT instead of Any-precision DNN | Pre-activation ResNet-20 · CIFAR-10, bitwidth (2,2) | accuracy: 93.97 -> 92.52 | Hurt |
| use MEBQAT instead of dedicated QAT at full precision | MobileNet-v2 · CIFAR-10, bitwidth (FP,FP) | accuracy: 93.00 -> 89.24 | Hurt |
| use MEBQAT-MAML instead of FOMAML | 5-layer CNN · Omniglot 20-way 1-shot, bitwidth (2,2) | few-shot accuracy: 25.97 -> 89.57 | Improved |
| use MEBQAT-MAML instead of FOMAML+QAT | 5-layer CNN · Omniglot 20-way 1-shot, bitwidth (2,2) | few-shot accuracy: 62.09 -> 89.57 | Improved |
| use MEBQAT-MAML instead of FOMAML+QAT | 5-layer CNN · MiniImageNet 5-way 5-shot, bitwidth (2,2) | few-shot accuracy: 62.25 -> 61.65 | No clear effect |
| use MEBQAT-PN instead of PN | 4-layer CNN · Omniglot 20-way 1-shot, bitwidth (2,2) | few-shot accuracy: 31.58 -> 94.87 | Improved |
| use MEBQAT-PN instead of PN+QAT | 4-layer CNN · MiniImageNet 5-way 1-shot, bitwidth (2,2) | few-shot accuracy: 50.06 -> 47.66 | Hurt |
| use MEBQAT-PN instead of PN at 3-bit | 4-layer CNN · MiniImageNet 5-way 5-shot, bitwidth (3,3) | few-shot accuracy: 46.74 -> 66.16 | Improved |
Closest papers by meaning
- Once Quantization-Aware Training: High Performance Extremely Low-Bit Architecture Search (ICCV 2021 · similarity 0.90)
- A Statistical Framework for Low-bitwidth Training of Deep Neural Networks (NeurIPS 2020 · similarity 0.90)
- MetaQuant: Learning to Quantize by Learning to Penetrate Non-differentiable Quantization (NeurIPS 2019 · similarity 0.89)
- Learning to Quantize for Training Vector-Quantized Networks (ICML 2025 · similarity 0.88)
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.87)
- PTMQ: Post-training Multi-Bit Quantization of Neural Networks (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.
