Performance improvements

[tavianator]

No description provided.

[tavianator]

I added some uses of C++11 that are breaking Travis. Can it be changed to pass -std=c++11, or do we need to maintain C++03 compatibility? I can also update the README and perhaps add a Makefile if you want.

[tavianator]

This post goes into some detail about the benefits/motivations of each patch: https://www.microsoft.com/en-us/research/blog/optimizing-barnes-hut-t-sne/

[lvdmaaten]

Thanks for submitting this and writing up your findings!

I think switching Travis to -std=c++11 should be fine.

For the vantage-tree point implementation, I have been meaning to switch to an approach based on product quantization for quite some time now, but I haven't had time to work on that yet.

I will do some local testing, and merge presuming the maps produced by the code still look as expected.

[tavianator]

Sounds great, thanks! I eyeballed the results on my test datasets before and after this series, and the differences were only in the least-significant couple of digits. I imagine they only differ at all due to the change in how I'm computing the centers of mass. If anything the new way should be slightly more accurate as it's not accumulating errors on every *(n-1)/n when new points are added.