Generalization errors. Dependency of the generalization error of the model from the number of features used and from the degree of non-linearity in the hidden layer. The values are average values over a 10-fold cross-validation procedure. Models with highest evidence (5 to 7 input features and 2 to 5 hidden neurons respectively) have a low generalization error below 0.25.
Schroeder et al. BMC Molecular Biology 2006 7:3 doi:10.1186/1471-2199-7-3