Research map / Architectures and efficiency / Attention and transformers
BiViT: Extremely Compressed Binary Vision Transformers
Yefei He, Zhenyu Lou, Luoming Zhang, Jing Liu, Weijia Wu, Hong Zhou, Bohan Zhuang
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 |
|---|---|---|---|
| replace BiBERT/BiT binarization with Softmax-aware Binarization, CLB and PWS | binary NesT-T (all binarized) · TinyImageNet | top-1 accuracy (%): 32.39 -> 52.21 | Improved |
| keep MLP activations full precision | binary NesT-T · TinyImageNet | top-1 accuracy (%): 52.21 -> 69.83 | Improved |
| add Softmax-aware Binarization to attention binarization (ablation) | BiBERT on NesT-T · TinyImageNet | top-1 accuracy (%): 68.51 -> 70.73 | Improved |
| use softmax in the backward gradient approximation (ablation) | SAB on NesT-T · TinyImageNet | top-1 accuracy (%): 68.71 -> 70.73 | Improved |
| use Cross-layer Binarization instead of one-step training (ablation) | binary NesT-T · TinyImageNet | top-1 accuracy (%): 58.20 -> 69.83 | Improved |
| use CLB instead of traditional two-step training (ablation) | binary NesT-T · TinyImageNet | top-1 accuracy (%): 65.64 -> 69.83 | Improved |
| add Parameterized Weight Scales to attention (ablation) | BiBERT on NesT-T · TinyImageNet | top-1 accuracy (%): 68.51 -> 72.75 | Improved |
| add Parameterized Weight Scales to MLP (ablation) | BiBERT on NesT-T · TinyImageNet | top-1 accuracy (%): 70.02 -> 71.35 | Improved |
| lower SAB threshold coefficient beta below 0.2 (ablation) | SAB on NesT-T · TinyImageNet | top-1 accuracy (%): 70.73 -> 69.18 | Hurt |
| vary beta between 0.25 and 0.45 (ablation) | SAB on NesT-T · TinyImageNet | top-1 accuracy (%): 70.73 / 70.81 / 70.68 | No clear effect |
| use BiT learned fixed thresholds instead of BiBERT Bool | binary Swin-T / NesT-T · TinyImageNet | top-1 accuracy (%): Swin-T 41.89 -> 40.52; NesT-T 32.39 -> 34.72 | Mixed |
| use local-window attention (Swin/NesT) instead of full attention (DeiT) | binary ViTs · ImageNet-1k val | top-1 accuracy (%): DeiT-T 37.9 vs Swin-T 70.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- BiBERT — BiBERT: Accurate Fully Binarized BERT (ICLR 2022) · better on 10 of 10 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 2024 · similarity 0.92)
- BVT-IMA: Binary Vision Transformer with Information-Modified Attention (AAAI 2024 · similarity 0.89)
- BiMLP: Compact Binary Architectures for Vision Multi-Layer Perceptrons (NeurIPS 2022 · similarity 0.88)
- MaxViT: Multi-axis Vision Transformer (ECCV 2022 · similarity 0.88)
- Learning Efficient Binarized Object Detectors With Information Compression (TPAMI 2021 · similarity 0.88)
- BinaryAttention: One-Bit QK-Attention for Vision and Diffusion Transformers (CVPR 2026 · similarity 0.88)
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.
