Research map / Architectures and efficiency / Quantization
Binarized Neural Machine Translation
Yichi Zhang, Ankush Garg, Yuan Cao, Lukasz Lew, Behrooz Ghorbani, Zhiru Zhang, Orhan Firat
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 |
|---|---|---|---|
| Binarize all dense-layer weights | 6L6L Transformer · WMT17 En-De | validation loss: 1.39 -> 1.38 | No clear effect |
| Binarize all dense-layer weights | 6L6L Transformer · WMT17 En-De | BLEU: 26.35 -> 25.93 | Hurt |
| Binarize FFN weights and activations | 6L6L Transformer · WMT17 En-De | validation loss: 1.39 -> 1.40 | No clear effect |
| Add FFN activation binarization to the all-dense-weight binary model (ablation) | 6L6L Transformer · WMT17 En-De | validation loss: 1.38 -> 1.51 | Hurt |
| Binarize QKV activations (ablation) | 1-bit dense-weight and FFN-activation model · WMT17 En-De | validation loss: 1.51 -> 1.72 | Hurt |
| Binarize attention output-projection activations (ablation) | 1-bit dense-weight and FFN-activation model · WMT17 En-De | validation loss: 1.51 -> 1.60 | Hurt |
| Binarize all attention dense-layer activations (ablation) | 1-bit dense-weight and FFN-activation model · WMT17 En-De | validation loss: 1.51 -> 1.89 | Hurt |
| Binarize QK and Score-V attention Einsum activations (ablation) | 1-bit dense-weight model · WMT17 En-De | BLEU: 25.93 -> 9.42 | Hurt |
| Use a scaling factor of at least 64 for binarized FFNs (ablation) | FFN-binarized Transformer · WMT17 En-De | training convergence: s=1; s=64 | Improved |
| Replace fixed scaling with LayerNorm (ablation) | Binarized attention output projection · WMT17 En-De | loss | Improved |
| Add a shortcut around the attention output projection (ablation) | Binarized attention output projection · WMT17 En-De | loss | Improved |
| Scale weight-binarized models in encoder or decoder depth | Binary Transformer · Train Subset and OOD evaluation sets | scaling behavior: 0.18 vs 0.16; 0.31 vs 0.28 | No clear effect |
Closest papers by meaning
- PB-LLM: Partially Binarized Large Language Models (ICLR 2024 · similarity 0.88)
- Mixture of Scales: Memory-Efficient Token-Adaptive Binarization for Large Language Models (NeurIPS 2024 · similarity 0.87)
- ARB-LLM: Alternating Refined Binarizations for Large Language Models (ICLR 2025 · similarity 0.87)
- BiLLM: Pushing the Limit of Post-Training Quantization for LLMs (ICML 2024 · similarity 0.86)
- ScaleBITS: Scalable Bitwidth Search for Hardware-Aligned Mixed-Precision LLMs (NeurIPS 2026 · similarity 0.85)
- Structured Transforms for Low-Overhead Quantization of Language Models (NeurIPS 2026 · similarity 0.85)
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.
