Learning with Transformation Invariant Kernels
Christian Walder1 and Olivier Chapelle2      Poster M40

  Gaussian kernel SVM:
TWO PARAMETERS:  length scale & noise parameter 

  Thinplate kernel SVM: (new)
ONE PARAMETER:  length scale = noise parameter  The joy of invariance  only one free parameter! They perform equally well in practice! We also derive basic results for  invariant and conditionally positive definite kernels.
 Max Planck Institute, Tübingen, Germany. 2  Yahoo! Research, Santa Clara, CA.
1