GeometricOptimizers

Documentation for GeometricOptimizers.

The package optimizes on homogeneous spaces — the Stiefel and Grassmann manifolds — with the standard neural network optimizers, by keeping the optimizer cache in a global tangent space that does not depend on the current iterate. Optimization on Homogeneous Spaces is the place to start; the Manifolds chapter builds up the geometry it rests on, and Optimizer Methods covers gradient descent, momentum and Adam themselves.

Neural networks are not here. GeometricMachineLearning builds them on top of this package and re-exports most of what the API Reference lists.