Research map / Architectures and efficiency / Quantization
Towards Effective Low-Bitwidth Convolutional Neural Networks
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 |
|---|---|---|---|
| apply PQ+TS+Guided instead of the baseline | AlexNet · ImageNet validation, 4-bit | top-1 accuracy: 56.8% -> 58.0% | Helped |
| apply PQ+TS+Guided instead of the baseline | AlexNet · ImageNet validation, 2-bit | top-1 accuracy: 48.8% -> 51.6% | Helped |
| apply PQ+TS+Guided instead of the baseline | ResNet-50 · ImageNet validation, 4-bit | top-1 accuracy: 75.1% -> 75.7% | Helped |
| apply PQ+TS+Guided instead of the baseline | ResNet-50 · ImageNet validation, 2-bit | top-1 accuracy: 67.7% -> 70.0% | Helped |
| apply PQ+TS+Guided instead of the baseline | AlexNet · CIFAR-100 test, 4-bit | top-1 accuracy: 65.0% -> 65.8% | Helped |
| apply PQ+TS+Guided instead of the baseline | AlexNet · CIFAR-100 test, 2-bit | top-1 accuracy: 63.9% -> 64.6% | Helped |
| add two-stage optimization (ablation) | 4-bit AlexNet trained from scratch with TS · CIFAR-100 | accuracy: 2.7 % | Helped |
| fine-tune from a full-precision model rather than learn from scratch (ablation) | low-bit AlexNet trained with TS · CIFAR-100 | validation accuracy | Helped |
| add guided training to PQ+TS (ablation) | AlexNet and ResNet-50 · ImageNet validation, 4-bit and 2-bit | top-1 accuracy: 57.8% -> 58.0%; 50.9% -> 51.6%; 75.5% -> 75.7%; 69.4% -> 70.0% | Helped |
| keep the first convolution and last fully connected layers full precision (ablation) | PQ+TS+Guided · ImageNet validation, AlexNet and ResNet-50, 4-bit and 2-bit | top-1 accuracy: 58.0% -> 58.1%; 51.6% -> 52.5%; 75.7% -> 75.9%; 70.0% -> 70.8% | Helped |
| jointly train the full-precision guide instead of keeping it unchanged (ablation) | Stage2+Guided AlexNet · ImageNet validation, 2-bit | top-1 accuracy | Helped |
Closest papers by meaning
- LEARNED STEP SIZE QUANTIZATION (ICLR 2020 · similarity 0.89)
- Incremental Network Quantization: Towards Lossless CNNs with Low-precision Weights (ICLR 2017 · similarity 0.89)
- Not All Bits have Equal Value: Heterogeneous Precisions via Trainable Noise (NeurIPS 2022 · similarity 0.89)
- Efficient Multi-bit Quantization Network Training via Weight Bias Correction and Bit-wise Coreset Sampling (NeurIPS 2025 · similarity 0.87)
- Towards Cheaper Inference in Deep Networks with Lower Bit-Width Accumulators (ICLR 2024 · similarity 0.87)
- AMPA: Adaptive Mixed Precision Allocation for Low-Bit Integer Training (ICML 2024 · similarity 0.86)
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.