Research map / Architectures and efficiency / RNNs and normalization
BiBench: Benchmarking and Analyzing Network Binarization
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 soft/custom gradient approximation (ReActNet, ReCU, FDA-style) instead of plain STE | binarized operators · Learning task and architecture tracks | OM_task / OM_arch | Helped |
| Binarize attention-based transformers (BERT-Tiny/Base) instead of CNNs/MLPs | binarized networks · Neural architecture track | OM_arch relative accuracy | Hurt |
| Use channel-wise statistical scaling factors vs. no scaling factor (BNN) | binarized operators · Learning task track | OM_task | Helped |
| Adopt Adam optimizer with 1x learning rate and CosineAnnealingLR for binarized training | binarized networks · Training consumption track (ResNet18, CIFAR10) | hyperparameter sensitivity | Helped |
| Use custom gradient approximation (soft approximation) in training | binarized networks (Bi-Real, ReActNet, FDA) · Training consumption track | training time cost | Hurt |
| Binarize weights and activations of ResNet-style CNN with pre-binarization mean-shifting (ReCU/ReActNet redistribution) | binarized networks · Learning task track | OM_task | Helped |
| Deploy binarized models on lower-power edge chips rather than high-performance chips | binarized models on ARM hardware · Hardware inference track | inference speedup | Helped |
| Add activation re-scaling (XNOR) or per-spatial scaling factors (XNOR++) to operators | deployment on Larq/daBNN · Hardware inference track | deployability | Hurt |
| Binarize ReActNet-style activations with mean-shifting (vs. none) | ReActNet deployment · Hardware inference track | OM_infer: ReActNet 11.51 vs BNN 11.70 | No clear effect |
| Binarize weights and activations with the original binary (sign) weights versus 1-bit quantization in CIFAR10-C | binarized ResNet · CIFAR10-C corruption robustness | relative robustness | No clear effect |
Closest papers by meaning
- PB-LLM: Partially Binarized Large Language Models (ICLR 2024 · similarity 0.86)
- BiMaCoSR: Binary One-Step Diffusion Model Leveraging Flexible Matrix Compression for Real Super-Resolution (ICML 2025 · similarity 0.86)
- STBLLM: Breaking the 1-Bit Barrier with Structured Binary LLMs (ICLR 2025 · similarity 0.85)
- BiDM: Pushing the Limit of Quantization for Diffusion Models (NeurIPS 2024 · similarity 0.85)
- Efficient Multitask Dense Predictor via Binarization (CVPR 2024 · similarity 0.84)
- BinaryDM: Accurate Weight Binarization for Efficient Diffusion Models (ICLR 2025 · 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.