Accelerating Natural Gradient Descent for PINNs with Randomized Numerical Linear Algebra

Name: Accelerating Natural Gradient Descent for PINNs with Randomized Numerical Linear Algebra
Uploaded: 2026-10-03T07:44:49.302751+00:00
Duration: 413 s
Description: This work improves Natural Gradient Descent (NGD) for training Physics-Informed Neural Networks (PINNs) by using Randomized Numerical Linear Algebra (RandNLA). The authors propose two methods, NyströmNGD and RPCholNGD, to efficiently precondition the inner solver. This results in significant speedup and better accuracy compared to standard NGD methods.

This work improves Natural Gradient Descent (NGD) for training Physics-Informed Neural Networks (PINNs) by using Randomized Numerical Linear Algebra (RandNLA). The authors propose two methods, NyströmNGD and RPCholNGD, to efficiently precondition the inner solver. This results in significant speedup and better accuracy compared to standard NGD methods.

Open the full episode