Research map / Architectures and efficiency / Quantization
Rethinking Differentiable Search for Mixed-Precision Neural Networks
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 |
|---|---|---|---|
| Use learned filter-wise mixed precision instead of uniform 2-bit quantization | AlexNet · ImageNet | top-1 accuracy: 58.6 -> 59.1 | Helped |
| Use learned filter-wise mixed precision instead of uniform 2-bit quantization | ResNet-18 · ImageNet | top-1 accuracy: 65.1 -> 65.9 | Helped |
| Use learned filter-wise mixed precision instead of uniform 2-bit quantization | GoogLeNet · ImageNet | top-1 accuracy: 64.8 -> 67.8 | Helped |
| Use learned filter-wise mixed precision instead of uniform 2-bit quantization | ResNet-50 · ImageNet | top-1 accuracy: 70.6 -> 72.1 | Helped |
| Use learned filter-wise mixed precision instead of uniform 2-bit quantization | Inception-V3 · ImageNet | top-1 accuracy: 71.0 -> 72.4 | Helped |
| Discretize the relaxed architecture using winner-take-all (ablation) | EdMIPS on ResNet-18 · ImageNet | classification accuracy | No clear effect |
| Share one weight tensor across bit-width candidates instead of learning separate tensors (ablation) | EdMIPS on ResNet-18 · ImageNet | classification accuracy | No clear effect |
| Sample architectures with multinomial sampling rather than categorical sampling (ablation) | EdMIPS on ResNet-18 · ImageNet | classification accuracy | Helped |
| Use single-pass optimization rather than alternating optimization (ablation) | EdMIPS on ResNet-18 · ImageNet | classification accuracy | No clear effect |
| Search a coarser bit-width space: weights {1, 4}, activations {2, 4} (ablation) | EdMIPS on ResNet-18 · ImageNet | classification accuracy | No clear effect |
| Replace parallel convolutions with an efficient composite convolution (ablation) | EdMIPS · ResNet-18 search | computation: 30-50% | Helped |
| Replace parallel convolutions with an efficient composite convolution (ablation) | EdMIPS · ResNet-18 search | memory: 20-40% | Helped |
Closest papers by meaning
- MetaMix: Meta-State Precision Searcher for Mixed-Precision Activation Quantization (AAAI 2024 · similarity 0.87)
- One-Shot Model for Mixed-Precision Quantization (CVPR 2023 · similarity 0.86)
- Mixed Precision DNNs: All you need is a good parametrization (ICLR 2020 · similarity 0.86)
- Towards Effective Low-Bitwidth Convolutional Neural Networks (CVPR 2018 · similarity 0.85)
- Learning from Loss Landscape: Generalizable Mixed-Precision Quantization via Adaptive Sharpness-Aware Gradient Aligning (ICML 2025 · similarity 0.85)
- AMPA: Adaptive Mixed Precision Allocation for Low-Bit Integer Training (ICML 2024 · similarity 0.83)
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.