[Master Index] [Index for export/lyngby]



Function Synopsis

[S, S1, Y, V, W, EAcc, Info] = lyngby_nns_main(X, T, arg1, arg2, ...

Help text

 lyngby_nns_main      - Main function for Neural network saliency

	function [S, S1, Y, V, W, EAcc, Info] = lyngby_nns_main(X, T,
	    'PropertyName', 'PropertyValue')

       Input:  X   The datamatrix
               T   Target output (the paradigm)
                  'Type'          [ {LS} | LSSecondOrder | 
                                  Entropy | EntropySecondOrder ]
                  'SVDComponents'   { 50 } Number of singular values
                                  maintained, ie. the number of
                                  neural network inputs.
                  'NNInput'       [ {SVD} | SVDNormalized | Direct |
                                  SOP | SOPNormalized | SOPResSVD ]
                                  Type of signal on the neural
                                  network input
                  'HiddenUnits'   { 3 } Number of hidden units, not
                                  counting the threshold unit
                  'Reg'           { 0.001 } Regularization parameter
                                   (weight decay)
                  'GenOptim'      [ {Free} | EarlyStop |
                                  HiddenUnitsEarlyStop | Pruning |
                                  Pruning1DRegGridSearch |
                                  Pruning2DReggridSearch ]
                                  Generalization optimization
                  'Validation'    [ {InBasisSingleBlocked} | 
                                  OffBasisSingleBlocked ]
                                  Type of validation (the type of
                                  validation set)
                  'Run'           Run specification, used in the
                                  automatic splitting of training and
                                  validation set. 
                  'Info'          [ {0} | 1 ] Continuous information
                                  about the optimization

       Output: S      Saliency
               S1     First order saliency
               E      Error for each example (each scan)
               V      Input weights
               W      Output weights
               EAcc   Is the accumulated error (the evolution of the
                      error), and it will depend on the setting of
               Info   Information about EAcc

       This function will first SVD the 'X' matrix into 'SVDComponents'
       singular values and rotations. The output from the SVD will be
       feed to a two-layer feed-forward neural network. This neural
       network will be optimized for the most generalizing prediction
       of 'T' (the paradigm) according to the setting of 'GenOptim'
       and 'Validation'. After this optimization the saliency of the
       variables in 'X' will be calculated.

       See also: lyngby_svd, lyngby_nn_emain, lyngby_nn_qmain

Cross-Reference Information

This function calls
This function is called by

Produced by mat2html on Wed Jul 29 15:43:40 2009
Cross-Directory links are: OFF