get_prep_info_class_mlp — Compute the information content of the preprocessed feature vectors of a multilayer perceptron.
get_prep_info_class_mlp computes the information content of the training vectors that have been transformed with the preprocessing given by Preprocessing. Preprocessing can be set to 'principal_components' or 'canonical_variates'. The preprocessing methods are described with create_class_mlp. The information content is derived from the variations of the transformed components of the feature vector, i.e., it is computed solely based on the training data, independent of any error rate on the training data. The information content is computed for all relevant components of the transformed feature vectors (NumInput for 'principal_components' and min(NumOutput - 1, NumInput) for 'canonical_variates', see create_class_mlp), and is returned in InformationCont as a number between 0 and 1. To convert the information content into a percentage, it simply needs to be multiplied by 100. The cumulative information content of the first n components is returned in the n-th component of CumInformationCont, i.e., CumInformationCont contains the sums of the first n elements of InformationCont. To use get_prep_info_class_mlp, a sufficient number of samples must be added to the multilayer perceptron (MLP) given by MLPHandle by using add_sample_class_mlp or read_samples_class_mlp.
InformationCont and CumInformationCont can be used to decide how many components of the transformed feature vectors contain relevant information. An often used criterion is to require that the transformed data must represent x% (e.g., 90%) of the data. This can be decided easily from the first value of CumInformationCont that lies above x%. The number thus obtained can be used as the value for NumComponents in a new call to create_class_mlp. The call to get_prep_info_class_mlp already requires the creation of an MLP, and hence the setting of NumComponents in create_class_mlp to an initial value. However, if get_prep_info_class_mlp is called it is typically not known how many components are relevant, and hence how to set NumComponents in this call. Therefore, the following two-step approach should typically be used to select NumComponents: In a first step, an MLP with the maximum number for NumComponents is created (NumInput for 'principal_components' and min(NumOutput - 1, NumInput) for 'canonical_variates'). Then, the training samples are added to the MLP and are saved in a file using write_samples_class_mlp. Subsequently, get_prep_info_class_mlp is used to determine the information content of the components, and with this NumComponents. After this, a new MLP with the desired number of components is created, and the training samples are read with read_samples_class_mlp. Finally, the MLP is trained with train_class_mlp.
Type of preprocessing used to transform the feature vectors.
Default value: 'principal_components'
List of values: 'principal_components', 'canonical_variates'
Relative information content of the transformed feature vectors.
Cumulative information content of the transformed feature vectors.
* Create the initial MLP create_class_mlp (NIn, NHidden, NOut, 'softmax', 'principal_components',\ NIn, 42, MLPHandle) * Generate and add the training data for J := 0 to NData-1 by 1 * Generate training features and classes * Data = [...] * Class = [...] add_sample_class_mlp (MLPHandle, Data, Class) endfor write_samples_class_mlp (MLPHandle, 'samples.mtf') * Compute the information content of the transformed features get_prep_info_class_mlp (MLPHandle, 'principal_components',\ InformationCont, CumInformationCont) * Determine NComp by inspecting InformationCont and CumInformationCont * NComp = [...] clear_class_mlp (MLPHandle) * Create the actual MLP create_class_mlp (NIn, NHidden, NOut, 'softmax', 'principal_components',\ NComp, 42, MLPHandle) * Train the MLP read_samples_class_mlp (MLPHandle, 'samples.mtf') train_class_mlp (MLPHandle, 100, 1, 0.01, Error, ErrorLog) write_class_mlp (MLPHandle, 'classifier.mlp') clear_class_mlp (MLPHandle)
If the parameters are valid, the operator get_prep_info_class_mlp returns the value 2 (H_MSG_TRUE). If necessary an exception is raised.
get_prep_info_class_mlp may return the error 9211 (Matrix is not positive definite) if Preprocessing = 'canonical_variates' is used. This typically indicates that not enough training samples have been stored for each class.
Christopher M. Bishop: “Neural Networks for Pattern Recognition”;
Oxford University Press, Oxford; 1995.
Andrew Webb: “Statistical Pattern Recognition”; Arnold, London; 1999.