perClass Documentation
version 5.4 (7-Dec-2018)
 SDNEURAL  Feed-forward or RBF neural network


   DATA      training dataset

  'MLP'  default Multi-Layer Perceptron
   'units',N     number of units in hidden layer (def: 10)
   'iters',N     number of iterations/epochs (def: 1000)
   'rate',R      training rate (def: 0.3)
   'momentum',M  momentum term (def: 0.8)
   'test',TS     external test set used for validation (def: [] = split DATA)
   'tsfrac',N    fraction of DATA used for validation (def: 0.2)
   'testeach',N  test each N iterations (def: 100)
   'targets',MAT target matrix with the same number of samples as DATA
   'noscale'     do not perform default data scaling
   'init',P      initialize from neural net P and continue training
  'RBF'  Radial basis function network
   'units',N     Number of units per class (def: 5)
   'partial'     Output pipeline returns outputs per hidden unit, not class output

   'gpu'         train on GPU (requires parallel computing toolbox)

   P         neural network pipeline
   RES       structure describing training process
    ETR,ETS   mean square error on training set and on test set

 SDNEURAL implements training of feed-forward (MLP) neural network with
 one hidden layer or Radial Basis Function (RBF) network.
 For the default feed-forward network, mean square error (MSE) is
 minimized with respect to the validation set (by default 20% of DATA).
 By default, input data set is scaled (can be disabled by 'noscale'
 option).  It is possible to continue training starting from existing
 neural net pipeline P with 'init' option. Use 'noscale' and, if needed,
 perform scaling manually using SDSCALE.
 SDNEURAL may approximate the user-supplied target matrix using the
 'targets' option.

 RBF network is trained using 'rbf' option. The number of units may be
 directly provided after 'rbf' option or specified with 'units'
 option. Default value is 5 units per class. A vector with num.of units
 per class may be specified. This is useful as simple unimodal classes
 need less units than multi-modal complex classes.
 The RBF network is trained with a direct formulation that does not
 perform iterative training. Also, there is no internal splitting of data
 as for MLP network. The 'partial' option returns the per-unit soft
 outputs instead of decisions.



sdneural is referenced in examples: