GOPHERSPACE.DE - P H O X Y
gophering on hngopher.com
HN Gopher Feed (2017-07-25) - page 1 of 10
 
___________________________________________________________________
Lessons Learned from Benchmarking Fast Machine Learning Algorithms
47 points by hoaphumanoid
https://blogs.technet.microsoft.com/machinelearning/2017/07/25/l...
7/07/25/lessons-learned-benchmarking-fast-machine-learning-algorithms/
___________________________________________________________________
 
nl - 1 hours ago
It's interesting that LightGBM was initially promoted as being more
accurate than XGB, but that claim always seemed marginal at best
and was hard to reproduce.Other investigations show the same thing
about training speed though, eg
https://medium.com/implodinggradients/benchmarking-lightgbm-...
 
matt4077 - 1 hours ago
The GPU versions are performing surprisingly bad. To even match CPU
performance, you need a training set in the tens of millions, and
even far beyond that, a doubling of speed seems to be the best you
can hope for.Compare to, for example, tensorflow, where it isn't
uncommon to see a 10x speedup even for moderately-sized training
sets.(I say "surprising" in the sense that I'm surprised; I don't
know the algorithms used for decision trees, and it may well be
that they are less amendable to GPU-parallelization than the NN-
and matrix algorithms I've worked with)