Research map / Architectures and efficiency / CNN design
Quantized Convolutional Neural Networks for Mobile Devices
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 response-error correction during quantization (ablation) | Q-CNN on AlexNet conv2 · ILSVRC-12 val, top-5 error increase | top-5 error increase: 8.97% -> 0.27% | Helped |
| Add response-error correction during quantization (ablation) | Q-CNN on MNIST 3-layer network · MNIST test | classification error: 1.42% -> 1.39% | Helped |
| Add response-error correction during quantization (ablation) | Q-CNN on MNIST 5-layer network · MNIST test | classification error: 1.34% -> 1.19% | Helped |
| Fine-tune subsequent layers after quantizing AlexNet conv2 (ablation) | Q-CNN (EC), AlexNet conv2 · ILSVRC-12 val | top-5 error increase: 0.27% -> 0.17% | Helped |
| Increase quantization precision from 4/64 to 6/64 (ablation) | Q-CNN (EC), AlexNet conv2 · ILSVRC-12 val | speed-up: 3.70× -> 5.36× | Helped |
| Increase quantization precision from 4/64 to 6/64 (ablation) | Q-CNN (EC), AlexNet conv2 · ILSVRC-12 val | top-5 error increase: 0.27% -> 0.50% | Hurt |
| Add response-error correction to convolutional-layer quantization (ablation) | Q-CNN on AlexNet, all convolutional layers · ILSVRC-12 val | top-5 error increase: 18.05% -> 0.50% | Helped |
| Add response-error correction to fully-connected quantization (ablation) | Q-CNN on CaffeNet first fully connected layer · ILSVRC-12 val | top-1 error increase: 0.35% -> 0.18% | Helped |
| Add response-error correction to fully-connected quantization (ablation) | Q-CNN on all CaffeNet fully connected layers · ILSVRC-12 val | top-5 error increase: 0.29% -> 0.30%; 0.47% -> 0.47%; 0.34% -> 0.27%; 0.59% -> 0.39% | Mixed |
| Quantize all convolutional layers with response-error correction | VGG-16 · ILSVRC-12 val | top-5 error increase: 0.95% -> 1.83% without fine-tuning; 0.35% -> 0.45% after fine-tuning | Mixed |
| Quantize the whole network | AlexNet · ILSVRC-12 val | top-5 error increase: 0.84% | Hurt |
Closest papers by meaning
- And the Bit Goes Down: Revisiting the Quantization of Neural Networks (ICLR 2020 · similarity 0.88)
- Fully Quantized Network for Object Detection (CVPR 2019 · similarity 0.88)
- Incremental Network Quantization: Towards Lossless CNNs with Low-precision Weights (ICLR 2017 · similarity 0.88)
- Focused Quantization for Sparse CNNs (NeurIPS 2019 · similarity 0.87)
- Multi-Precision Quantized Neural Networks via Encoding Decomposition of {-1,+1} (AAAI 2019 · similarity 0.87)
- GradQ-ViT: Robust and Efficient Gradient Quantization for Vision Transformers (AAAI 2025 · similarity 0.85)
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.