You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
KD-tree / Ball tree, but we use ellipses (e.g. covariances) which may only contain 68% of points,
Metric tree / M-tree, but we dont always need the triangle inequality (or may not have it on Riemannian manifolds),
Working on-manifold also implies that we have to mix point and tangent definitions (e.g. Gaussian-type / Radial Basis kernel covariances).
Much of the involved computations will occur in any number of tangent spaces given all the points on the manifold. "Tangent" also amplifies the notion of curved-spaces by not assuming Euclidean. Also, tangent has a relevant word play with Jacobian, which is heavily used here for optimization, path discovery and more.
Kernel is chosen since the tree represents recursively decomposing kernels. I.e. computations against the root kernel should be similar as doing computations against any number of tree or leaf kernels. Kernel replaces Ball, since the metrics/distances/divergences that will be used are then particular to that instantiation of the tree.
Legacy name
Is an amalgamation from {Manifold + Elliptical + Metric} Tree, but no one will know what that is. Name change is intended to try streamline recognition with more established (but no longer fitting) names.
Alternate candidates
The following names were also considered
StatsBallTree
ManifoldMetricTree
The text was updated successfully, but these errors were encountered:
New name
Derives from
Working on-manifold also implies that we have to mix point and tangent definitions (e.g. Gaussian-type / Radial Basis kernel covariances).
Much of the involved computations will occur in any number of tangent spaces given all the points on the manifold. "Tangent" also amplifies the notion of curved-spaces by not assuming Euclidean. Also, tangent has a relevant word play with Jacobian, which is heavily used here for optimization, path discovery and more.
Kernel is chosen since the tree represents recursively decomposing kernels. I.e. computations against the root kernel should be similar as doing computations against any number of tree or leaf kernels. Kernel replaces Ball, since the metrics/distances/divergences that will be used are then particular to that instantiation of the tree.
Legacy name
Is an amalgamation from {Manifold + Elliptical + Metric} Tree, but no one will know what that is. Name change is intended to try streamline recognition with more established (but no longer fitting) names.
Alternate candidates
The following names were also considered
The text was updated successfully, but these errors were encountered: