Research map / Architectures and efficiency / Quantization
Differentiable Joint Pruning and Quantization for Hardware Efficiency
Ying Wang, Yadong Lu, Tijmen Blankevoort
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 |
|---|---|---|---|
| optimize pruning and quantization jointly instead of in two stages (prune then quantize) (ablation) | DJPQ on ResNet18 · ImageNet | top-1 accuracy (%): 68.52 -> 69.27 | Improved |
| optimize pruning and quantization jointly instead of in two stages (prune then quantize) (ablation) | DJPQ on ResNet18 · ImageNet | BOP compression ratio: 46.53 -> 52.94 | Improved |
| add joint pruning to mixed-precision quantization (DJPQ vs DQ) | mixed-precision quantization on ResNet18 · ImageNet | BOPs (G): 40.71G vs 30.87G BOPs | Improved |
| add joint pruning to mixed-precision quantization (DJPQ vs DQ) | mixed-precision quantization on MobileNetV2 · ImageNet | BOPs (G): 175.24G vs 132.28G BOPs | Improved |
| learn power-of-two bit-widths during training instead of rounding DQ bit-widths afterwards | mixed-precision quantization on ResNet18 · ImageNet | BOP compression ratio: 31.59 -> 52.28 | Improved |
| naively round DQ learned bit-widths up to powers of two | DQ on ResNet18 · ImageNet | BOP compression ratio: 45.53 -> 31.59 | Hurt |
| restrict bit-widths to powers of two within DJPQ | DJPQ on VGG7 · CIFAR10 | top-1 accuracy (%): 91.54 -> 91.43 | No clear effect |
| restrict bit-widths to powers of two within DJPQ | DJPQ on ResNet18 · ImageNet | top-1 accuracy (%): 69.27 -> 69.12 | No clear effect |
| use ADAM instead of SGD with manually scaled learning rates | DJPQ training · ImageNet | accuracy | Hurt |
Closest papers by meaning
- OPQ: Compressing Deep Neural Networks with One-shot Pruning-Quantization (AAAI 2021 · similarity 0.89)
- Single-Path Bit Sharing for Automatic Loss-Aware Model Compression (TPAMI 2023 · similarity 0.89)
- Automatic Joint Structured Pruning and Quantization for Efficient Neural Network Training and Compression (CVPR 2025 · similarity 0.88)
- JointSQ: Joint Sparsification-Quantization for Distributed Learning (CVPR 2024 · similarity 0.88)
- JAQ: Joint Efficient Architecture Design and Low-Bit Quantization with Hardware-Software Co-Exploration (AAAI 2025 · similarity 0.87)
- Effective Interplay between Sparsity and Quantization: From Theory to Practice (ICLR 2025 · similarity 0.87)
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.
