Research map / Architectures and efficiency / Quantization
Ultra-Low Precision 4-bit Training of 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 |
|---|---|---|---|
| use radix-4 FP4 gradient format instead of radix-2 (ablation) | 4-bit training · ResNet18 CIFAR10 | accuracy | Helped |
| add per-layer trainable GradScale instead of APEX global scaling (ablation) | FP4 training · ResNet18 ImageNet | accuracy: > 1% vs. APEX | Helped |
| add two-phase rounding (TPR) (ablation) | FP4 backward and update GEMMs · ResNet18 ImageNet | test accuracy: by >3.5% | Helped |
| compute bottom 1x1 conv layers in FP8 (ablation) | TPR FP4 training · ResNet50 ImageNet | accuracy: +1.0% | Helped |
| compute both 1x1 conv layers in FP8 (ablation) | TPR FP4 training · ResNet50 ImageNet | accuracy: +1.5% | Helped |
| full 4-bit training instead of FP32 | ResNet50 · ImageNet | accuracy: 2.5% degradation | Hurt |
| use FP8 for update GEMM (dW) (ablation) | TPR FP4 LSTM · SWB300 | WER: ~1.4% -> < 0.5% loss | Helped |
| swap even/odd phase assignment in TPR (ablation) | TPR FP4 · ResNet18 ImageNet | convergence | No clear effect |
Closest papers by meaning
- Accurate Neural Training with 4-bit Matrix Multiplications at Standard Formats (ICLR 2023 · similarity 0.91)
- Optimizing Large Language Model Training Using FP4 Quantization (ICML 2025 · similarity 0.87)
- AMPA: Adaptive Mixed Precision Allocation for Low-Bit Integer Training (ICML 2024 · similarity 0.87)
- ScaleBITS: Scalable Bitwidth Search for Hardware-Aligned Mixed-Precision LLMs (NeurIPS 2026 · similarity 0.87)
- SDP4Bit: Toward 4-bit Communication Quantization in Sharded Data Parallelism for LLM Training (NeurIPS 2024 · similarity 0.87)
- Normalized Architectures are Natively 4-Bit (NeurIPS 2026 · 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.