Research map / Probabilistic and scientific modeling / Differential equations and PDEs
STDE++: Polynomial-Time Amortization for Linear Differential Operators
Zekun Shi, Zheyuan Hu, Min Lin, Kenji Kawaguchi
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 |
|---|---|---|---|
| use JAX instead of PyTorch for stacked backward-mode SDGD | Backward mode SDGD · Two-body Allen-Cahn, 100 D | training speed (it/s): 55.56 -> 40.63 (JAX slower at 100 D); about 15x speed-up overall | Improved |
| parallelize the sampled-dimension loop of SDGD via HVP (ablation) | Backward mode SDGD (JAX) · Two-body Allen-Cahn, 10K D | training speed (it/s): 29.85 -> 216.83 | Improved |
| use mixed-mode forward-over-backward AD instead of stacked backward (ablation) | Parallelized backward mode SDGD · Two-body Allen-Cahn, 10K D | training speed (it/s): 216.83 vs 193.91 | No clear effect |
| replace parallelized SDGD with STDE | Parallelized backward mode SDGD · Two-body Allen-Cahn, 100K D | training speed (it/s): 29.24 -> 156.90 | Improved |
| replace parallelized SDGD with STDE | Parallelized backward mode SDGD · Two-body Allen-Cahn, 100K D | peak memory (MB): 4931 -> 1073 | Improved |
| replace Forward Laplacian with STDE at high dimension | Forward Laplacian · Two-body Allen-Cahn, 10K D | training speed (it/s): 32.15 -> 454.16 | Improved |
| replace Forward Laplacian with STDE at low dimension | Forward Laplacian · Two-body Allen-Cahn, 100 D | training speed (it/s): 1974.50 -> 1035.09 | Hurt |
| replace original PyTorch SDGD with STDE | Backward mode SDGD (PyTorch) · Two-body Allen-Cahn, 100K D | peak memory (MB): 32777 -> 1073 | Improved |
| replace original PyTorch SDGD with STDE | Backward mode SDGD (PyTorch) · Two-body Allen-Cahn, 100K D | training speed (it/s): 0.23 -> 156.90 | Improved |
| use STDE to scale to 1M dimensions | STDE · Two-body Allen-Cahn, 1M D | training speed (it/s): OOM -> 13.61 | Improved |
Later papers that compared with this method
1 later accepted paper reports STDE as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Overcoming PINNs Failure Modes In High Dimension With Low-Rank Fourier Sum (ICML 2026 · better on 1 of 1 shared result)
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
- Stochastic Taylor Derivative Estimator: Efficient amortization for arbitrary differential operators (NeurIPS 2024 · similarity 0.96)
- Amortized Fourier Neural Operators (NeurIPS 2024 · similarity 0.86)
- Derivative-enhanced Deep Operator Network (NeurIPS 2024 · similarity 0.86)
- Efficient Training of Neural Fractional-Order Differential Equation via Adjoint Backpropagation (AAAI 2025 · similarity 0.85)
- Derivative-Informed Training of Neural Operators On-the-Fly via Sketched Tangent Consistency (NeurIPS 2026 · similarity 0.85)
- Optimal deep learning of holomorphic operators between Banach spaces (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.
