Research map / Architectures and efficiency / Quantization
Incremental Network Quantization: Towards Lossless CNNs with Low-precision Weights
Read the paperCheck 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 |
|---|---|---|---|
| Incrementally quantize weights to 5-bit powers-of-two-or-zero values with partition and retraining | AlexNet, VGG-16, GoogleNet, ResNet-18, and ResNet-50 · ImageNet validation center crop | top-1 error: 0.15%/2.28%/0.13%/0.71%/1.59% decrease | Helped |
| Incrementally quantize weights to 5-bit powers-of-two-or-zero values with partition and retraining | AlexNet, VGG-16, GoogleNet, ResNet-18, and ResNet-50 · ImageNet validation center crop | top-5 error: 0.23%/1.65%/0.25%/0.41%/1.21% decrease | Helped |
| Replace random partition with pruning-inspired weight partition (ablation) | INQ on ResNet-18 · ImageNet validation center crop | top-1 error: 32.11% -> 31.02%; 1.09% decrease | Helped |
| Replace random partition with pruning-inspired weight partition (ablation) | INQ on ResNet-18 · ImageNet validation center crop | top-5 error: 11.73% -> 10.90%; 0.83% decrease | Helped |
| Reduce expected bit-width from 5-bit to 4-bit (ablation) | INQ on ResNet-18 · ImageNet validation center crop | top-1 error: 31.02% -> 31.11% | No clear effect |
| Reduce expected bit-width from 5-bit to 4-bit (ablation) | INQ on ResNet-18 · ImageNet validation center crop | top-5 error: 10.90% -> 10.99% | No clear effect |
| Reduce expected bit-width to 3-bit (ablation) | INQ on ResNet-18 · ImageNet validation center crop | top-1 error: 31.02% -> 31.92%; 0.19% loss versus full-precision reference | Hurt |
| Reduce expected bit-width to 3-bit (ablation) | INQ on ResNet-18 · ImageNet validation center crop | top-5 error: 10.90% -> 11.64%; 0.33% loss versus full-precision reference | Hurt |
| Reduce expected bit-width to 2-bit ternary weights (ablation) | INQ on ResNet-18 · ImageNet validation center crop | top-1 error: 31.73% -> 33.98%; 2.25% decrease in top-1 error rate | Hurt |
| Reduce expected bit-width to 2-bit ternary weights (ablation) | INQ on ResNet-18 · ImageNet validation center crop | top-5 error: 11.31% -> 12.87%; 1.56% decrease in top-5 error rate | Hurt |
| Combine pruning with INQ rather than the reported deep-compression approach | AlexNet · ImageNet AlexNet | compression ratio: 53$\times$ | Helped |
| Add 4-bit weight quantization to the pruning-plus-INQ compression setup (ablation) | AlexNet · ImageNet AlexNet | compression ratio: 71$\times$ | Helped |
Closest papers by meaning
- Explicit Loss-Error-Aware Quantization for Low-Bit Deep Neural Networks (CVPR 2018 · similarity 0.92)
- LEARNED STEP SIZE QUANTIZATION (ICLR 2020 · similarity 0.90)
- Learning to Quantize Deep Networks by Optimizing Quantization Intervals With Task Loss (CVPR 2019 · similarity 0.90)
- And the Bit Goes Down: Revisiting the Quantization of Neural Networks (ICLR 2020 · similarity 0.89)
- Post training 4-bit quantization of convolutional networks for rapid-deployment (NeurIPS 2019 · similarity 0.89)
- SYQ: Learning Symmetric Quantization for Efficient Deep Neural Networks (CVPR 2018 · 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.