Research map / Architectures and efficiency / Attention and transformers
Visual Dependency Transformers: Dependency Tree Emerges From Reversed Attention
Mingyu Ding, Yikang Shen, Lijie Fan, Zhenfang Chen, Zitian Chen, Ping Luo, Joshua B. Tenenbaum, Chuang Gan
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 |
|---|---|---|---|
| Reversed attention (dependency tree induction) in weakly-supervised pretraining | DependencyViT-T vs DeiT-Tiny · Pascal-Part mIoU | mIoU (%): 18.9 -> 23.2 with maximum spanning | Improved |
| Reversed attention with self-supervised pretraining (vs iBOT baseline) | DependencyViT-T · Pascal-Part mIoU | mIoU (%): 25.1 -> 28.7 | Improved |
| Reversed attention with self-supervised pretraining (vs iBOT baseline) | DependencyViT-T · Car-Parts mIoU | mIoU (%): 25.7 -> 27.0 | Improved |
| Soft dependency mask added to features for saliency | DependencyViT-T (weakly-sup.) · ECSSD maxF_beta | maxF_beta (%): DeiT 49.3 vs DependencyViT 62.0 | Improved |
| Dependency-based model vs DeiT on unsupervised saliency (weakly-sup.) | DependencyViT-T · DUTS maxF_beta | maxF_beta (%): 34.2 -> 53.8 | Improved |
| Self-supervised instead of weakly-supervised pretraining for saliency | DependencyViT-T · DUT-OMRON maxF_beta | maxF_beta (%): 52.0 (weakly) vs 32.5 (self) | Hurt |
| Dynamic pruning of leaf nodes (DependencyViT-Lite) | DependencyViT-T ImageNet classification · ImageNet-1K Top-1 | Top-1 (%): 75.4 -> 73.7 with 0.8 vs 1.3 GFLOPs | Hurt |
| Reversed attention (full DependencyViT) vs standard ViT | Tiny-scale ImageNet classification · ImageNet-1K Top-1 | Top-1 (%): 72.2 -> 75.4 | Improved |
| Pruning fewer tokens (keep 128 vs 64) (ablation) | DependencyViT-Lite-T · ImageNet-1K Top-1 | Top-1 (%): 73.7 (64 tokens) -> 74.9 (128 tokens) | Improved |
| Head selector and message controller (P, M) added to reversed attention (ablation) | DependencyViT-T · ablation on tiny models | accuracy/ablation score: not reported numerically in provided text | Improved |
Closest papers by meaning
- Linear Differential Vision Transformer: Learning Visual Contrasts via Pairwise Differentials (NeurIPS 2025 · similarity 0.86)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.85)
- Causal Tracing of Object Representations in Large Vision Language Models: Mechanistic Interpretability and Hallucination Mitigation (AAAI 2026 · similarity 0.85)
- Automated Detection of Visual Attribute Reliance with a Self-Reflective Agent (NeurIPS 2025 · similarity 0.85)
- Register and [CLS] tokens induce a decoupling of local and global features in large ViTs (NeurIPS 2025 · similarity 0.85)
- To Sink or Not to Sink: Visual Information Pathways in Large Vision-Language Models (ICLR 2026 · 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.
