April 20, 2024
Fisher Flow: Optimization on the Statistical Manifold
Gradient descent in Euclidean space ignores the geometry of probability distributions. Natural gradient descent uses the Fisher information metric instead. Fisher Flow makes this continuous.