Research map / Architectures and efficiency / Quantization
HLHLp: Quantized Neural Networks Training for Reaching Flat Minima in Loss Surface
Sungho Shin, Jinhwan Park, Yoonho Boo, Wonyong Sung
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 |
|---|---|---|---|
| replace fine-tuning with HLHLp training | 2-bit ResNet-14 · CIFAR-10 | test accuracy: 89.20 -> 90.64 | Improved |
| replace fine-tuning with HLHLp training | 2-bit ResNet-32 · CIFAR-10 | test accuracy: 92.32 -> 93.05 | Improved |
| replace fine-tuning with HLHLp training | 2-bit ResNet-20 · CIFAR-100 | test accuracy: 64.47 -> 66.44 | Improved |
| replace fine-tuning with HLHLp training | 2-bit MobileNetV2 · CIFAR-100 | test accuracy: 74.97 -> 75.51 | Improved |
| add HLHLp second stage vs HL only on contaminated ImageNet (ablation) | 2-bit ResNet-18 · ImageNet-C noise | average noise accuracy: 39.08% -> 45.99% | Improved |
| quantize PTB GRU with HLHLp instead of Wang et al. ternary | 2-bit GRU · WikiText-2 | test PPL: 132.49 -> 105.96 | Improved |
| quantize LSTM with HLHLp 4-level instead of Xu et al. | 2-bit LSTM · WikiText-2 | test PPL: 106.10 -> 105.5 | Improved |
| use HLHLp for CTC acoustic model instead of Lee et al. | 2-bit LSTM AM · WSJ test | CER: 9.76 -> 8.21 | Improved |
| use HLHLp for CTC acoustic model instead of Lee et al. | 2-bit LSTM AM · Aurora-4 noisy | CER: 51.65 -> 48.6 | Improved |
| ablation (B): float cyclic learning-rate training for 110 epochs then 2-bit (ablation) | GRU PTB LM · PTB | PPL: 96.97 -> 115.92 | Hurt |
| ablation (C): replace 8-bit high-precision step with 2-bit (ablation) | GRU PTB LM · PTB | PPL: 96.97 -> 97.84 | Hurt |
| ablation (D): monotonic decreasing learning rate fine-tuning only (ablation) | GRU PTB LM · PTB | PPL: 96.97 -> 112.08 | Hurt |
Closest papers by meaning
- Training Quantized Nets: A Deeper Understanding (NeurIPS 2017 · similarity 0.89)
- Explicit Loss-Error-Aware Quantization for Low-Bit Deep Neural Networks (CVPR 2018 · similarity 0.88)
- Quantized Neural Networks: Training Neural Networks with Low Precision Weights and Activations (JMLR 2018 · similarity 0.88)
- Fine-tuning Quantized Neural Networks with Zeroth-order Optimization (ICLR 2026 · similarity 0.87)
- Up or Down? Adaptive Rounding for Post-Training Quantization (ICML 2020 · similarity 0.87)
- PikeLPN: Mitigating Overlooked Inefficiencies of Low-Precision Neural Networks (CVPR 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.
