NGD Geometric Bias Paper

A research paper published on arXiv that systematically analyzes the geometric distance between the Adam optimizer and natural gradient descent. It decomposes the Adam update rule as a diagonal empirical Fisher approximation and quantifies directional deviation across multiple loss surfaces using a scale-invariant metric.