Research map / Architectures and efficiency / Quantization
Training Quantized Neural Networks With a Full-Precision Auxiliary Module
Check 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 |
|---|---|---|---|
| Add a full-precision auxiliary module to 2-bit DoReFa-Net | ResNet-50 · ImageNet validation | Top-1 accuracy: 70.2 -> 73.8 | Helped |
| Add a full-precision auxiliary module to 2-bit DoReFa-Net | ResNet-101 · ImageNet validation | Top-1 accuracy: 70.8 -> 74.6 | Helped |
| Add classification losses at intermediate layers | 2-bit DoReFa-Net, ResNet-18 · ImageNet validation | Top-1 accuracy: 64.7 -> 64.9 | No clear effect |
| Replace auxiliary learning with knowledge distillation | 2-bit DoReFa-Net, ResNet-50 · ImageNet validation | Top-1 accuracy: 73.8 -> 71.4 | Hurt |
| Remove skip connections and add Auxi | DoReFa-Net plain ResNet-34 · ImageNet validation | Top-1 accuracy: 62.1 -> 66.4 | Helped |
| Add Auxi to the plain quantized network | DoReFa-Net plain ResNet-18 · CIFAR-100 test | Top-1 accuracy: 64.6 -> 67.9 | Helped |
| Use a 3×3 adaptor instead of a 1×1 adaptor (ablation) | DoReFa-Net, ResNet-18 · ImageNet validation | Top-1 accuracy: 66.7 -> 66.9 | Helped |
| Add auxiliary learning (ablation) | 4-bit RetinaNet, ResNet-18, shared heads · COCO minival | AP: 29.2 -> 30.6 | Helped |
| Do not share prediction heads across feature pyramid levels (ablation) | 4-bit RetinaNet with Auxi, ResNet-18 · COCO minival | AP: 30.6 -> 31.9 | Helped |
| Do not share prediction heads across feature pyramid levels (ablation) | Full-precision RetinaNet, ResNet-18 · COCO minival | AP: 32.1 -> 31.2 | Hurt |
| Quantize the feature pyramid and prediction heads as well as the backbone (ablation) | 4-bit RetinaNet, ResNet-18 · COCO minival | AP: 32.1 -> 29.2 | Hurt |
Closest papers by meaning
- Quantization Networks (CVPR 2019 · similarity 0.88)
- FracBits: Mixed Precision Quantization via Fractional Bit-Widths (AAAI 2021 · similarity 0.88)
- Towards Effective Low-Bitwidth Convolutional Neural Networks (CVPR 2018 · similarity 0.88)
- Post-training Quantization with Multiple Points: Mixed Precision without Mixed Precision (AAAI 2021 · similarity 0.88)
- AMPA: Adaptive Mixed Precision Allocation for Low-Bit Integer Training (ICML 2024 · 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.