If data dimension is much larger than the requested number of components k, linear PCA pre-processing of the data is recommended for reducing the dimension of the data before applying nonlinear PCA. This can automatically be done by using the option 'pre_pca'.
Example
data dimension: 20 (or larger)
requested components: k=3
The dimension of the input data will be reduced to 12 by setting this number as required input (and output) units of the nonlinear network structure.
[pc,net,network]=nlpca(data,3, 'pre_pca','yes', 'units_per_layer',[12, 8, 3, 8, 12] );