Here are
some applications that you can freely download. They come with no guarantee!
1) Mono-Concave Multilayer Perceptron Builder
(uses genetic algorithms to train neural
networks with specified signs for 1st and 2nd order
derivatives of the output with respect to some of its inputs)
MONMLP-guide.pdf
(a short user's guide)
(A java user friendly version of the algorithm. Requires Java
1.4). After unzipping the MONMLP.zip file you may run the application by executing
the file monmlp.bat (Windows only). In a Linux
environment you may launch (from the directory
where the application was unzipped) the following command: java -classpath jfreechart-0.9.4.jar:jcommon-0.7.1.jar:.: Interfata
For more
details on how this algorithm works please consult the second chapter of the
thesis.
2) Pattern
recognition functions tools
The package tarcafun contains several useful functions
implemented in MATLAB® which applies to problems of pattern recognition.
Download first tarcafun.zip and unzip it. Set your Matlab
path to point to the directory where you unzipped the archive. Now you may use
the following functions:
|
Function Name |
Summary Description |
Help |
|
featureSelector |
Is a feature selection algorithm for CLASSIFICATION and REGRESSION. The relevance criterion is J=alpha*AR+gama*ADC where Accuracy Rate (AR) is obtained by crossvalidation with a k-nearest neighbor classifier
(CLASSIFICATION only) and ADC is the Asymmetric Dependency Coefficient from
information theoretic framework. The combinatorial optimization method is
"plus l take away r". This function
indicates which among the columns of a matrix X are best suited to map y by
ranking them in the decreasing order of their importance. |
|
|
featureSelectorInit |
This function
is almost similar with the function featureSelector
(see help for details) with the exception that it may start with a initial guess of features which are relevant. |
|
|
itss_ADC |
Computes
an Asymmetric Dependency Coefficient (Information theory) between a feature
variable X and a response variable y. |
|
|
lau_confmatrix |
Returns the confusion
matrix for a classifier. |
|
|
lau_knn |
Implements
a k-nearest neighbor algorithm. This is a simple classification rule which
assigns for a new point x the class in which are found most of the k nearest
neighbors of x in the training data set. |
|
|
lau_knn_crosserr |
Computes the performance
of a k-nearest neighbor classifier by n-fold crossvalidation.
The global misclassification rate, the confusion matrix and its standard
deviation are computed. At each fold, a disjoint fraction of data (1/cvsets)
is predicted while the remaining fraction (1-1/cvsets) is used as prototypes.
The confusion matrix is computed for each fold and summed up to form the
global confusion matrix, gen. |
|
|
weight_eval |
Interprets the weight
matrices of a trained feed-forward neural network. Gives as return the
importance (saliency) index for each of the inputs of the neural network.
After an idea by Garson G. D.(1991) extended for
multi output networks. |