Research map / Architectures and efficiency / Quantization
DSConv: Efficient Convolution Operator
Marcelo Gennari do Nascimento, Roger Fawcett, Victor Adrian Prisacariu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| reduce block size from 256 to 8 for 3-bit weights (ablation) | DSConv GoogLeNet · ImageNet val, weight-only | top-1 accuracy: 5.7 -> 56.8 | Improved |
| reduce block size from 256 to 4 for 2-bit weights (ablation) | DSConv VGG19 · ImageNet val, weight-only | top-1 accuracy: 11.3 -> 69.4 | Improved |
| reduce block size from 256 to 4 for 3-bit weights (ablation) | DSConv ResNet50 · ImageNet val, weight-only | top-1 accuracy: 44.6 -> 74.7 | Improved |
| reduce block size from 256 to 4 for 4-bit weights (ablation) | DSConv ResNet50 · ImageNet val, weight-only | top-1 accuracy: 73.0 -> 75.6 | Improved |
| add distillation adaptation with unlabelled data (3/3 bits, B=64) (ablation) | DSConv ResNet50 · ImageNet val | top-1 accuracy: 54.4 -> 71.5 | Improved |
| add distillation adaptation with unlabelled data (3/3 bits, B=64) (ablation) | DSConv GoogLeNet · ImageNet val | top-1 accuracy: 29.3 -> 56.7 | Improved |
| add distillation adaptation with unlabelled data (3/3 bits, B=64) (ablation) | DSConv ResNet18 · ImageNet val | top-1 accuracy: 30.1 -> 65.8 | Improved |
| add distillation adaptation with unlabelled data (4/4 bits, B=128) (ablation) | DSConv ResNet50 · ImageNet val | top-1 accuracy: 72.6 -> 74.2 | Improved |
| lower weight bit width from 4 to 2 without retraining (ablation) | DSConv GoogLeNet · ImageNet val, no data | top-1 accuracy: 63.3 -> 27.6 | Hurt |
| lower weight bit width from 4 to 2 without retraining (ablation) | DSConv ResNet50 · ImageNet val, no data | top-1 accuracy: 75.1 -> 65.1 | Hurt |
| minimise KL divergence instead of L2 norm for the shift values (ablation) | DSConv weight quantization · ImageNet val | top-1 accuracy | No clear effect |
| lower activation and weight bits from 8/8 to 4/4 without adaptation (ablation) | DSConv ResNet18 · ImageNet val, no data | top-1 accuracy: 69.7 -> 67.3 | Hurt |
Closest papers by meaning
- Post training 4-bit quantization of convolutional networks for rapid-deployment (NeurIPS 2019 · similarity 0.89)
- And the Bit Goes Down: Revisiting the Quantization of Neural Networks (ICLR 2020 · similarity 0.88)
- Dynamic Network Quantization for Efficient Video Inference (ICCV 2021 · similarity 0.88)
- Focused Quantization for Sparse CNNs (NeurIPS 2019 · similarity 0.87)
- CondConv: Conditionally Parameterized Convolutions for Efficient Inference (NeurIPS 2019 · similarity 0.87)
- Effective Interplay between Sparsity and Quantization: From Theory to Practice (ICLR 2025 · 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.
