Understanding PINN’s Failure Modes || Progressive Sharpening from Jacobian Alignment|| Nov 14, 2025
Keywords
Summary
110 words
Critical Evaluation
Value of the Information & Strength of the Argument
The talks provide valuable insights into PINN failure modes, challenging the common belief that they are due to bad local minima. The first talk offers a detailed analysis and a novel architecture (PINN Mamba) with experimental results. The second talk presents a clear explanation for progressive sharpening with supporting experiments. However, the argumentation is somewhat weakened by the lack of comprehensive comparison with existing methods (e.g., causality-based approaches) and the use of absolute errors instead of relative errors in some results.
Scientific Rigor, Source Quality, Title Accuracy
The speakers reference their own published papers (ICML, NeurIPS) and open-source code, which adds credibility. The title accurately reflects the content. However, the seminar does not provide a full list of sources, and the discussion reveals potential gaps in baseline comparisons. The Q&A highlights concerns about the claim of state-of-the-art performance without thorough comparison.
150 words
Title / Content Match
The title accurately reflects the content, which covers two talks on PINN failure modes and progressive sharpening.
Quality & Reliability
7/10
The seminar presents two research talks with technical depth and references to published papers (ICML, NeurIPS). However, the discussion includes critical questions about baseline comparisons and the claim of state-of-the-art, indicating some limitations in the presented results.
Key Moments
Markers derived by PSI from the transcript: the creator did not define chapters.
- Introduction and speaker introductions
- First talk begins: PINN failure modes overview
- Discussion of continuous-discrete mismatch and simplicity bias
- Introduction of PINN Mamba architecture and results
- Q&A session on baseline comparisons
- Second talk begins: Progressive sharpening and Jacobian alignment
- Experimental results and power law scaling
- Conclusion and final remarks
Cited Sources
- PINN Mamba (ICML 2025) — Mentioned as open-source project
- NeurIPS 2025 paper on PINN failure modes — Mentioned as published work
Concurring Sources
- PINN Mamba (ICML 2025) — Open-source implementation mentioned in the talk
Dissenting Sources
- Causality-based PINN methods — During Q&A, a participant noted that causality-based methods can solve the same problems with high accuracy, questioning the claim of state-of-the-art without comparison.
Contribution & Novelties
The seminar provides original contributions: a new understanding of PINN failure modes as a transient phase rather than a local minimum, and a novel architecture (PINN Mamba) that addresses propagation issues. The second talk offers a mechanistic explanation for progressive sharpening via Jacobian alignment. These insights could guide future research in PINN optimization and training dynamics.
Pour aller plus loin :
- Physics-Informed Neural Networks — Overview of PINNs.
- State Space Models — Background on state space models.
- Jacobian matrix — Mathematical concept relevant to the second talk.
87 words
Radar Profile
The radar profile shows high scores in quantity of information and technical level, with moderate scores in quality and reliability. This indicates a technically dense seminar with valuable insights, but with some limitations in source rigor and comparison.
💬 No comments were provided for this video.