Au to m a ti c online tuning for fast Gaussian summation (Poster W61)
Vlad I. Morariu, Balaji V. Srinivasan, Vikas C. Raykar, Ramani Du raiswam i, Larry S. Davis · Co st is quadratic · S l o ws down kernel machines

Previous method (IFGT) takes linear time but
· i s optimized for uniform distributions · p e rfo rm s poorly for small bandwidths

O u r Solution:

· u s e tree data structure

· a u to m a ti ca l l y tu n e to distribution · fo u r methods, each optimal in different situations · co m p a re s favorably with Dual-Tree methods as well

Bl a ck box approach: Automatically predict fastest method and t u n e its parameters for the given dataset.
Wa n t to speed up your code? Download FIGTree ( o p e n source):

http://sourceforge.net/projects/figtree