Research map / Architectures and efficiency / Quantization
Any-Precision Deep Neural Networks
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 |
|---|---|---|---|
| Train a shared model jointly across bit-widths, with separate BatchNorm states for each bit-width | Resnet-50 · ImageNet | accuracy: 63.18 vs 61.08 | Helped |
| Train a shared model jointly across bit-widths, with separate BatchNorm states for each bit-width | Resnet-20 · Cifar-10 | accuracy: 93.80 vs 93.66 | No clear effect |
| Train a shared model jointly across bit-widths, with separate BatchNorm states for each bit-width | CNN-8 · SVHN | accuracy: 88.21 vs 90.94 | Hurt |
| Add knowledge distillation from higher-precision outputs (ablation) | Resnet-20 · Cifar-10 | accuracy: 91.67 -> 92.15 | Helped |
| Add knowledge distillation from higher-precision outputs (ablation) | Resnet-50 · ImageNet | accuracy: 62.42 -> 63.18 | Helped |
| Train with candidate bit-widths 1,2,4,8 rather than 1,8 (ablation) | Resnet-20 · Cifar-10, tested at 3 bit | accuracy: 93.22 vs 92.97 | Helped |
| Train with candidate bit-widths 1,8 rather than 2,8 (ablation) | Resnet-20 · Cifar-10, tested at 1 bit | accuracy: 91.95 vs 10.00 | Helped |
| Train with candidate bit-widths 1,8 rather than 4,8 (ablation) | Resnet-20 · Cifar-10, tested at 2 bit | accuracy: 89.57 vs 72.78 | Helped |
| Add BatchNorm calibration after bit-shifting a full-precision model | Resnet-50 · ImageNet, 4 bit | accuracy: 42.70 -> 57.12 | Helped |
| Add BatchNorm calibration after bit-shifting a full-precision model | Resnet-50 · ImageNet, 1 bit | accuracy: 0.100 -> 0.106 | No clear effect |
| Replace fixed-precision post-training quantization with the trained any-precision model | Resnet-50 · ImageNet, 1 bit | accuracy: 0.100 -> 63.18 | Helped |
| Replace fixed-precision post-training quantization with the trained any-precision model | Resnet-50 · ImageNet, 4 bit | accuracy: 71.64 -> 74.75 | Helped |
Closest papers by meaning
- Not All Bits have Equal Value: Heterogeneous Precisions via Trainable Noise (NeurIPS 2022 · similarity 0.88)
- Mixed Precision Training (ICLR 2018 · similarity 0.88)
- Any-Precision LLM: Low-Cost Deployment of Multiple, Different-Sized LLMs (ICML 2024 · similarity 0.86)
- Towards Cheaper Inference in Deep Networks with Lower Bit-Width Accumulators (ICLR 2024 · similarity 0.85)
- The Effect of Weight Precision on the Neuron Count in Deep ReLU Networks (ICML 2024 · similarity 0.85)
- PikeLPN: Mitigating Overlooked Inefficiencies of Low-Precision Neural Networks (CVPR 2024 · similarity 0.84)
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.