From Non-Convex Self-Concordant Regularization to Scalable Quasi-Newton Training of PINNs
The episode discusses 'SCORE,' a new quasi-Newton optimizer for Physics-Informed Neural Networks (PINNs). The authors adapted self-concordance theory, traditionally used for convex problems, to handle the complex, non-convex optimization landscape of PINNs. SCORE achieves significantly higher accuracy across multiple challenging PDEs by improving the optimization process without altering the underlying physics objective.