ANaGRAM: A Natural Gradient Relative to Adapted Metrics for efficient PINNs learning
Date:
Presentation of ANaGRAM, a theoretically founded and algorithmically efficient natural gradient method for training Physics-Informed Neural Networks (PINNs). The talk covers the functional analysis perspective on PINNs training, the connection to the Neural Tangent Kernel, and an efficient SVD-based implementation that reduces the complexity of natural gradient descent significantly.