Research map / Architectures and efficiency / Attention and transformers
Choose a Transformer: Fourier or Galerkin
Shuhao Cao
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 softmax attention with Galerkin-type attention | operator learners for Burgers and Darcy solution operators · Burgers, n=8192 | relative error (× 10^-3): 4.151 -> 1.025 | Improved |
| Use an attention-based operator learner instead of FNO | Darcy flow solution operator · Darcy flow, n_f,n_c=141,43 | relative error (× 10^-2): 1.419 -> 0.838 | Improved |
| Use attention-based operator learners instead of FNO | inverse Darcy coefficient identification · Inverse Darcy, n_f,n_c=141,36, ε=0.1 | relative error (× 10^-2): 15.12 -> 6.659 | Improved |
| Use the Galerkin projection-type layer normalization rather than regular layer normalization (ablation) | GT on unnormalized Burgers data · Burgers, n=8192 | relative error (× 10^-3): 2.747 -> 1.025 | Improved |
| Use the Galerkin projection-type layer normalization rather than regular layer normalization (ablation) | GT on normalized Darcy data · Darcy flow, n_f,n_c=141,43 | relative error (× 10^-2): 0.894 -> 0.839 | No clear effect |
| Add diagonal initialization with η=δ=10^-2 (ablation) | GT with the new normalization on Burgers · Burgers, n=8192 | relative error (× 10^-3): 9.129 -> 1.025 | Improved |
| Add diagonal initialization with η=δ=10^-2 (ablation) | GT with regular normalization on Burgers · Burgers, n=512 | relative error (× 10^-3): 200.4 -> 15.14 | Improved |
| Add attention dropout ζ=0.1 (ablation) | GT with regular normalization on Burgers · Burgers, n=8192 | relative error (× 10^-3): 217.5 -> 207.0 | Improved |
| Remove the softmax normalization | attention-based operator learners · training efficiency | GFLOP per BP: 1393 -> 275 | Improved |
| Use Galerkin attention instead of linear softmax attention | encoder-only profiling, n=8192, d=128, l=10 · training speed | iterations per second: 10.98 -> 19.20 | Improved |
Closest papers by meaning
- Improved Operator Learning by Orthogonal Attention (ICML 2024 · similarity 0.89)
- Positional Knowledge is All You Need: Position-induced Transformer (PiT) for Operator Learning (ICML 2024 · similarity 0.87)
- Data-Efficient Operator Learning via Unsupervised Pretraining and In-Context Learning (NeurIPS 2024 · similarity 0.86)
- Optimal deep learning of holomorphic operators between Banach spaces (NeurIPS 2024 · similarity 0.86)
- Geometry Aware Operator Transformer as an efficient and accurate neural surrogate for PDEs on arbitrary domains (NeurIPS 2025 · similarity 0.86)
- Pretraining Codomain Attention Neural Operators for Solving Multiphysics PDEs (NeurIPS 2024 · similarity 0.85)
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.
