Research map / Architectures and efficiency / Quantization
Post training 4-bit quantization of convolutional networks for rapid-deployment
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 |
|---|---|---|---|
| Combine ACIQ activation clipping, per-channel bit allocation, and weight bias correction | 4-bit weights and 4-bit activations · ImageNet validation across VGG, VGG-BN, IncepV3, Res18, Res50, and Res101 | Top-1 accuracy: 67.2% -> 70.5%; 64.5% -> 71.8%; 30.6% -> 66.4%; 51.6% -> 67.0%; 62.0% -> 73.8%; 62.6% -> 75.0% | Helped |
| Add ACIQ activation clipping and per-channel bit allocation (ablation) | 8W4A activation quantization · ImageNet validation across VGG, VGG-BN, IncepV3, Res18, Res50, and Res101 | Top-1 accuracy: 68.8% -> 70.7%; 70.6% -> 72.8%; 70.9% -> 75.1%; 61.5% -> 68.0%; 68.3% -> 73.6%; 66.5% -> 75.6% | Helped |
| Add weight bias correction and per-channel bit allocation (ablation) | 4W8A weight quantization · ImageNet validation across VGG, VGG-BN, IncepV3, Res18, Res50, and Res101 | Top-1 accuracy: 70.5% -> 71.2%; 68.5% -> 72.4%; 38.4% -> 68.2%; 59.7% -> 68.3%; 72.5% -> 75.3%; 74.6% -> 76.9% | Helped |
| Use ACIQ activation clipping alone (ablation) | 8W4A activation quantization · ImageNet validation across six models | Top-1 accuracy: 68.8%, 70.6%, 70.9%, 61.5%, 68.3%, 66.5% -> 70.1%, 72.0%, 72.7%, 66.6%, 71.8%, 72.6% | Helped |
| Use per-channel bit allocation alone for activations (ablation) | 8W4A activation quantization · ImageNet validation across six models | Top-1 accuracy: 68.8%, 70.6%, 70.9%, 61.5%, 68.3%, 66.5% -> 69.7%, 72.6%, 74.3%, 65.0%, 71.3%, 70.8% | Helped |
| Use weight bias correction alone (ablation) | 4W8A weight quantization · ImageNet validation across six models | Top-1 accuracy: 70.5%, 68.5%, 38.4%, 59.7%, 72.5%, 74.6% -> 71.0%, 71.7%, 59.5%, 67.4%, 74.8%, 76.3% | Helped |
| Use per-channel bit allocation alone for weights (ablation) | 4W8A weight quantization · ImageNet validation across six models | Top-1 accuracy: 70.5%, 68.5%, 38.4%, 59.7%, 72.5%, 74.6% -> 71.0%, 71.9%, 61.4%, 66.7%, 75.0%, 76.4% | Helped |
| Clip weights with ACIQ (ablation) | The proposed quantization approach · Weight quantization | Top-1 accuracy | No clear effect |
Closest papers by meaning
- Accurate Neural Training with 4-bit Matrix Multiplications at Standard Formats (ICLR 2023 · similarity 0.92)
- Accurate Post Training Quantization With Small Calibration Sets (ICML 2021 · similarity 0.90)
- And the Bit Goes Down: Revisiting the Quantization of Neural Networks (ICLR 2020 · similarity 0.90)
- Incremental Network Quantization: Towards Lossless CNNs with Low-precision Weights (ICLR 2017 · similarity 0.89)
- Towards Unified INT8 Training for Convolutional Neural Network (CVPR 2020 · similarity 0.89)
- Efficient Multi-bit Quantization Network Training via Weight Bias Correction and Bit-wise Coreset Sampling (NeurIPS 2025 · 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.