Intrinsic Triangulation of Loss Landscapes of Neural Networks
Recent work investigating geometry and topology of loss landscapes of neural networks
revealed interesting properties regarding connectivity of clusters of local minima
[1],
[2].
In this work you will try to apply advances in computation intrinsic triangulation for 3d rendering
[3] towards those surfaces
(or reasonable hyperplane projections).
One possible motivation for this is the change for an ability to construct geodesic
paths on the constructed approximation.
Most (if not all) of the programming will be done in Python.
Inquiries: Tobias Hille