Poster
in
Workshop: Duality Principles for Modern Machine Learning
On the Fisher-Rao Gradient of the Evidence Lower Bound
Jesse van Oostrum · Nihat Ay
Keywords: [ variational autoencoder ] [ natural gradient ] [ information geometry ] [ evidence lower bound ]
This article studies the Fisher-Rao gradient, also referred to as the natural gradient, of the evidence lower bound, known from machine learning. Based on invariance properties of gradients within information geometry, it derives conditions on the underlying model that ensure the exactness for this bound. Information geometry is naturally based on duality concepts from differential geometry.