Research map / Label-efficient and robust learning / Recognition and long-tail
Consistency Driven Sequential Transformers Attention Model for Partially Observable Scenes
Check 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 |
|---|---|---|---|
| Add soft teacher-prediction consistency loss (ablation) | STAM · ImageNet, first two glimpses | accuracy: ~3% | Helped |
| Add soft teacher-prediction consistency loss (ablation) | STAM · fMoW, first two glimpses | accuracy: ~8% | Helped |
| Use hard rather than soft teacher pseudo-labels for consistency (ablation) | STAM · ImageNet, first two glimpses | accuracy gain over omitting consistency loss: ~1.5% | Helped |
| Use hard rather than soft teacher pseudo-labels for consistency (ablation) | STAM · fMoW, first two glimpses | accuracy gain over omitting consistency loss: ~3.5% | Helped |
| Replace the heuristic glimpse sequence with a learned attention policy | STAM versus the best fixed-order baseline · ImageNet, t=8 | accuracy: 1.8% | Helped |
| Replace the heuristic glimpse sequence with a learned attention policy | STAM versus the best fixed-order baseline · fMoW, t=8 | accuracy: 2.3% | Helped |
| Increase glimpse size from 32×32 to 48×48 or 64×64 (ablation) | STAM · ImageNet and fMoW, below 20% area observed | accuracy | Hurt |
| Increase the core capacity from DeiTᴰ-Tiny toward DeiTᴰ-Base (ablation) | STAM · ImageNet | accuracy | Helped |
| Train STAM for 400 rather than 200 epochs (ablation) | STAM · ImageNet, five glimpses | accuracy: 1.15% | Helped |
| Train STAM for 400 rather than 200 epochs (ablation) | STAM · fMoW | accuracy | Hurt |
| Stop sensing when predicted-class confidence exceeds 0.5 | STAM · ImageNet | accuracy: 66.47% | Helped |
| Stop sensing when predicted-class confidence exceeds 0.5 | STAM · fMoW | accuracy: 65.71% | Helped |
Closest papers by meaning
- LookHere: Vision Transformers with Directed Attention Generalize and Extrapolate (NeurIPS 2024 · similarity 0.85)
- XAttention: Block Sparse Attention with Antidiagonal Scoring (ICML 2025 · similarity 0.85)
- Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · similarity 0.85)
- Skip-Attention: Improving Vision Transformers by Paying Less Attention (ICLR 2024 · similarity 0.84)
- Breaking the Low-Rank Dilemma of Linear Attention (CVPR 2025 · similarity 0.84)
- Win-Win: Training High-Resolution Vision Transformers from Two Windows (ICLR 2024 · 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.