Research map / Architectures and efficiency / Quantization
Learning to Quantize Deep Networks by Optimizing Quantization Intervals With Task Loss
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 |
|---|---|---|---|
| Optimize trainable pruning and clipping intervals with task loss for weights and activations | QIL · ImageNet val, ResNet-18, 2/2 A/W | top-1 accuracy (%): 65.7 vs 64.9 | Helped |
| Progressively finetune from higher bit-width networks instead of directly finetuning from full precision (ablation) | QIL · ImageNet val, ResNet-18, 2/2 A/W | top-1 accuracy (%): 56.0 -> 65.7 | Helped |
| Jointly train network weights with quantizers instead of freezing weights (ablation) | QIL · ImageNet val, ResNet-18, 4/4 A/W | top-1 accuracy (%): 68.0 -> 70.1 | Helped |
| Jointly train network weights with quantizers instead of freezing weights (ablation) | QIL · ImageNet val, ResNet-18, 2/2 A/W | top-1 accuracy (%): 20.9 -> 65.7 | Helped |
| Train gamma rather than fixing it to 1.0 (ablation) | QIL · ImageNet val, AlexNet, 2/2 A/W | top-1 accuracy (%): 57.2 -> 58.1 | Helped |
| Train gamma rather than fixing it to 1.0 (ablation) | QIL · ImageNet val, AlexNet, 3/3 A/W | top-1 accuracy (%): 61.3 -> 61.4 | No clear effect |
| Train gamma rather than fixing it to 1.0 (ablation) | QIL · ImageNet val, ResNet-34, 2/2 A/W | top-1 accuracy (%): 70.6 -> 70.6 (+0.0) | No clear effect |
| Train gamma rather than fixing it to 1.0 (ablation) | QIL · ImageNet val, ResNet-18, 2/2 A/W | top-1 accuracy (%): 65.7 -> 66.1 (+0.4) | Helped |
| Fix gamma to 1.5 rather than train gamma (ablation) | QIL · ImageNet val, AlexNet, 2/2 A/W | top-1 accuracy (%): 58.1 -> 56.8 | Hurt |
| Quantize weights only while leaving activations at 32 bits | QIL · ImageNet val, ResNet-18, 32/2 A/W | top-1 accuracy (%): 68.1 vs 68.0 | Helped |
Closest papers by meaning
- LEARNED STEP SIZE QUANTIZATION (ICLR 2020 · similarity 0.90)
- Incremental Network Quantization: Towards Lossless CNNs with Low-precision Weights (ICLR 2017 · similarity 0.90)
- Explicit Loss-Error-Aware Quantization for Low-Bit Deep Neural Networks (CVPR 2018 · similarity 0.89)
- Network Quantization With Element-Wise Gradient Scaling (CVPR 2021 · similarity 0.89)
- Efficient Quantization for Neural Networks with Binary Weights and Low Bitwidth Activations (AAAI 2019 · similarity 0.88)
- Deep Neural Network Quantization via Layer-Wise Optimization Using Limited Training Data (AAAI 2019 · similarity 0.88)
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.