Research map / Architectures and efficiency / Quantization
Per-Tensor Fixed-Point Quantization of the Back-Propagation Algorithm
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 |
|---|---|---|---|
| Use per-tensor fixed-point precision configuration C_o instead of floating-point training | Back-propagation training · CIFAR-10 ConvNet | test error: 12.02% -> 12.58% | Hurt |
| Use per-tensor fixed-point precision configuration C_o instead of floating-point training | Back-propagation training · SVHN ConvNet | test error: 2.43% -> 2.58% | Hurt |
| Use per-tensor fixed-point precision configuration C_o instead of floating-point training | Back-propagation training · CIFAR-10 ResNet | test error: 7.42% -> 7.51% | Hurt |
| Use per-tensor fixed-point precision configuration C_o instead of floating-point training | Back-propagation training · CIFAR-100 ResNet | test error: 28.06% -> 27.43% | Helped |
| Increase every tensor precision by 1 bit above C_o (ablation) | FX (C_o) · CIFAR-10 ConvNet | test error | Helped |
| Increase every tensor precision by 1 bit above C_o (ablation) | FX (C_o) · Other evaluated networks | test error | No clear effect |
| Decrease every tensor precision by 1 bit below C_o (ablation) | FX (C_o) · Evaluated networks | test error | Hurt |
| Decrease weight precision (ablation) | FX (C_o) · Evaluated networks | test error | Hurt |
| Decrease activation-gradient precision (ablation) | FX (C_o) · Evaluated networks | test error | Hurt |
| Use stochastic quantization rather than the method's quantization | CIFAR-10 ConvNet · CIFAR-10 ConvNet | test error: 12.58% -> 11.32% | Helped |
| Use stochastic quantization rather than the method's quantization | SVHN ConvNet · SVHN ConvNet | test error: 2.58% -> 2.73% | Hurt |
| Use stochastic quantization rather than the method's quantization | CIFAR-10 ResNet · CIFAR-10 ResNet | test error: 7.51% -> 7.42% | No clear effect |
Closest papers by meaning
- Fixed-Point Back-Propagation Training (CVPR 2020 · similarity 0.89)
- Not All Bits have Equal Value: Heterogeneous Precisions via Trainable Noise (NeurIPS 2022 · similarity 0.88)
- One-Shot Model for Mixed-Precision Quantization (CVPR 2023 · similarity 0.87)
- Analytical Guarantees on Numerical Precision of Deep Neural Networks (ICML 2017 · similarity 0.86)
- LEARNED STEP SIZE QUANTIZATION (ICLR 2020 · similarity 0.86)
- Towards Cheaper Inference in Deep Networks with Lower Bit-Width Accumulators (ICLR 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.