Dominant Aligned Subspace Analysis on PINNs || Feature Tracking in PINNs || Oct 30, 2026

Value of the Information & Strength of the Argument

The first talk offers a strong conceptual contribution: it challenges the conventional dominance of sharpest Hessian directions by demonstrating that gradient-aligned subspaces are more predictive of optimization progress. Theoretical arguments are backed by illustrative toy models and empirical benchmarks, with clear comparisons to existing methods like Dom-SGD and Bulk-SGD. The line-search analysis and the observation that DAS can reduce the effect of stochastic noise in mini-batch training are particularly insightful. However, the method is explicitly a proof-of-concept; no computational complexity or scaling to large models is provided, and many claims are empirical rather than rigorously proven. The second talk, though only abstracted, proposes an elegant geometric approach (deformation maps) that naturally addresses the long-standing issue of localized feature resolution in PINNs, but its efficacy and limitations cannot be assessed from the transcript alone. Overall, the arguments are coherent and well-paced, with a clear logical structure, but the lack of exhaustive validation and the absence of external references weaken the argumentation’s completeness.

Scientific Rigor, Source Quality, Title Accuracy

The talks are scientifically rigorous in methodology: they involve full-batch Hessian computation, eigendecomposition, and careful experimental comparisons. The authors correctly identify numerical instabilities and address them via eigenvalue cutoff. However, no concrete sources (papers, datasets, code repositories) are cited during the talks, and the description provides no external links; mentions of prior work (e.g., ‘Cohen 2021’ for edge of stability, ‘Dom-SGD’, ‘Bulk-SGD’) are informal without precise bibliographic detail. This limits verifiability. The title matches the content accurately, covering both talks, and the channel is a legitimate academic seminar series. Overall, the rigor lies in the mathematical exposition and experimental design, but source quality is constrained by a lack of explicit citation.

The title accurately reflects the content, as the seminar consists of two distinct talks: one on dominant aligned subspace analysis and another on feature tracking in PINNs.

7/10

The seminar presents advanced mathematical and empirical findings on optimization subspaces for PINNs, with detailed theoretical framing and experimental results. However, no external sources are cited with URLs, and the methods are presented as proof-of-concept without computational efficiency analysis, limiting immediate verification.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:45Introduction of the seminar and speaker Ahmad AliAhmad.
• 02:30Discussion of edge of stability and prior work on low-dimensional subspaces.
• 08:00Definition of Dominant Aligned Subspace (DAS) and its construction.
• 15:00Experimental results on 1D Poisson/heat/wave equations showing DAS performance with line search.
• 21:40Analysis of edge of stability collapse and the effect of line search on preventing it.
• 30:00Mini-batch training results and the benefits of using a full-batch subspace.
• 40:00Gauss-Newton within DAS: results showing even smaller subspace requirements.
• 50:00Discussion of computational challenges and future directions.
• 55:00Q&A session begins after first talk.

• Edge of Stability (Cohen et al., 2021) — Mentioned as foundational work on the edge of stability phenomenon in neural network training.
• Dom-SGD and Bulk-SGD — Cited as prior low-dimensional subspace methods based on Hessian eigenvalue magnitude.