Olivier Chapelle

- preprocessing: all components were normalized to have variance = 1.
- feature selection: none
- classifier: SVM with L2 penalization of the slacks and RBF kernel. No transduction.
  Threshold adjusted afterwards to optimize an (approximate) leave-one-out BER.
- Hyperparameters (C and sigma) optimized by gradient descent on eiter: leave-one-out, radius/margin bound, validation error (in that case, half of the training set is used as validation), evidence (Bayesian treatment).
- Performance prediction based on an approximate leave-one-out procedure (no use of the test set).
 - paper: relevant material can be found at http://www.kyb.mpg.de/publication.html?publ=1436
 - code available at: http://www.kyb.tuebingen.mpg.de/bs/people/chapelle/ams/ams.m