mlpack  blog
Implementing probabilistic KDE error bounds - Week 5

Implementing probabilistic KDE error bounds - Week 5

Roberto Hueso, 03 July 2019

This week has been mostly about reclaiming not used probability. Probability reclaim has been implemented for both single and dual trees.

The idea is this: Whenever a KDE's ruleset evaluates kernel on two different points by any means that are not Monte Carlo estimation, there's some amount of probability that it's not being used. If we could use that amount of probability for making future Monte Carlo estimation less strict, then the original total amount of probability would still be the same and a higher amount of point's KDE can be estimated by using Monte Carlo, which reduces computation time.

Results might depend very much on the selected parameters but, for all tests I did, I got an increase on the amount of points estimated by Monte Carlo.

All code needs to be reviewed but in the meantime I'll be working on subspace trees.

Don't throw away stuff XKCD