Research map / Architectures and efficiency / Attention and transformers
SparseBERT: Rethinking the Importance Analysis in Self-attention
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 end-to-end differentiable attention-mask learning instead of manually designed sparse masks | BERT-base · GLUE development set | average: 80.1 -> 82.0 | Helped |
| Use structured DAM with λ=10^-1 instead of BigBird's manually designed mask | BERT-base · GLUE development set | average: 79.4 -> 79.6 | Helped |
| Increase λ to obtain a sparser unstructured attention mask | DAM_u · GLUE development set | average: 82.0 -> 80.8 | Hurt |
| Increase λ to obtain a sparser structured attention mask | DAM_s · GLUE development set | average: 80.9 -> 79.6 | Hurt |
| Remove diagonal attention (ablation) | Sparse attention masks · GLUE development set | GLUE: 79.5 -> 80.2; 79.7 -> 79.6; 80.1 -> 80.1; 77.9 -> 77.8; 79.4 -> 79.5; 78.9 -> 78.6; 80.5 -> 80.9; 79.3 -> 79.6 | No clear effect |
| Remove diagonal attention (ablation) | Sparse attention masks · GLUE development set | sparsity: 70.4 -> 71.2; 72.7 -> 73.4; 88.7 -> 89.5; 89.8 -> 90.6; 93.2 -> 93.9; 96.1 -> 96.9; 90.4 -> 91.2; 92.7 -> 93.5 | Helped |
| Train the mask end-to-end rather than thresholding attention probabilities in a separate stage (ablation) | DAM attention masks · QNLI and MRPC development sets | performance | Helped |
Closest papers by meaning
- Linear Log-Normal Attention with Unbiased Concentration (ICLR 2024 · similarity 0.89)
- SEA: Sparse Linear Attention with Estimated Attention Mask (ICLR 2024 · similarity 0.89)
- Low-Rank Approximation for Sparse Attention in Multi-Modal LLMs (CVPR 2024 · similarity 0.89)
- Sparse Attention as Compact Kernel Regression (NeurIPS 2026 · similarity 0.89)
- Transformers Learn Faster with Semantic Focus (NeurIPS 2025 · similarity 0.89)
- Theory, Analysis, and Best Practices for Sigmoid Self-Attention (ICLR 2025 · 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.