I always thought that ML is statistics + geometry done on the computer in high-dimensional spaces. Geometry comes in to help with high dimensionality. Instead of learning distributions exactly which is very hard in high dimensions we learn separating spaces of relatively simple geometry (like hyperplanes) as approximations.
I recently have been watching the old talk with lex… glad to see Michael did not change is mind
I always thought that ML is statistics + geometry done on the computer in high-dimensional spaces. Geometry comes in to help with high dimensionality. Instead of learning distributions exactly which is very hard in high dimensions we learn separating spaces of relatively simple geometry (like hyperplanes) as approximations.