Research map / Architectures and efficiency / Attention and transformers
ConvBERT: Improving BERT with Span-based Dynamic Convolution
Zi-Hang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan
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 conventional self-attention with ConvBERT mixed attention | BERTbase · GLUE test | Avg.: 80.9 -> 84.9 | Improved |
| use ConvBERTbase and train for 4M updates | BERTbase · GLUE test | Avg.: 83.5 -> 86.4 | Improved |
| add span-based dynamic convolution to bottleneck attention (ablation) | ConvBERTsmall · GLUE dev | GLUE: 80.6 -> 81.4 | Improved |
| add span-based dynamic convolution to bottleneck attention (ablation) | ConvBERTbase · GLUE dev | GLUE: 84.9 -> 85.7 | Improved |
| add grouped linear feed-forward operator (ablation) | ConvBERTmedium-small, BNK · GLUE dev | GLUE: 80.9 -> 81.0 | No clear effect |
| increase hidden and head dimensions (ablation) | ConvBERTmedium-small, BNK + GL · GLUE dev | GLUE: 81.0 -> 81.1 | No clear effect |
| replace span-based dynamic convolution with conventional convolution (ablation) | ConvBERTsmall · GLUE dev | Avg.: 81.4 -> 78.2 | Hurt |
| replace span-based dynamic convolution with token-conditioned dynamic convolution (ablation) | ConvBERTsmall · GLUE dev | Avg.: 81.4 -> 80.6 | Hurt |
| add conventional depth-wise separable convolution parallel to self-attention (ablation) | BERT · GLUE dev | Avg. | Hurt |
| add token-conditioned dynamic convolution (ablation) | BERT · GLUE dev | Avg. | No clear effect |
| use mixed attention instead of self-attention | base-sized model · inference | Speed (ms/sample): 17.66 -> 12.94 | 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.
- ELECTRAbase — ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators (ICLR 2020) · better on 4 of 5 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
- AutoBERT-Zero: Evolving BERT Backbone from Scratch (AAAI 2022 · similarity 0.85)
- Revisiting Few-sample BERT Fine-tuning (ICLR 2021 · similarity 0.85)
- The Heads Hypothesis: A Unifying Statistical Approach Towards Understanding Multi-Headed Attention in BERT (AAAI 2021 · similarity 0.85)
- Incorporating BERT into Parallel Sequence Decoding with Adapters (NeurIPS 2020 · similarity 0.85)
- Scaling Stick-Breaking Attention: An Efficient Implementation and In-depth Study (ICLR 2025 · similarity 0.84)
- Unveiling and Harnessing Hidden Attention Sinks: Enhancing Large Language Models without Training through Attention Calibration (ICML 2024 · similarity 0.83)
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.
