Demonstration by Nadia Vendrig: perClass - all you need to build classifiers
The perClass toolbox (previously known PRSD studio) was developed by Dr. Pavel Paclik and can be used to design classifiers in Matlab. Eighteen different classifiers (e.g. random forest, neural network, support vector machines) are included in the current version of perClass. It contains tools to visualize and understand the data, train and test classifiers, and optimize classifier performance. These classifiers can be exported and executed outside Matlab.
In perClass it is very easy to build a hierarchy of classifiers (cascade) and to visualize and adapt performance of the classifier. When separating apples and bananas for instance, we could first train a detector (e.g. Gaussian detector) on all data to protect from outliers, and in the second step train a classifier (e.g. mixture of Gaussians) to distinguish between apples and bananas. The result of this classification can be evaluated and adapted using ROC analysis, to better comply with the specific demands of the project. If bananas would be more expensive than apples, we could adapt the classifier such that the number of bananas misclassified as apples is reduced at the cost of error on apples.
Between the 3rd and the 7th of December, I took the PerClass-course in Delft and I was pleasantly surprised by its performance and user-friendliness. For those of you who are interested, I would like to give a demonstration of the toolbox so you can see what the software can do and how it works.
More information on the PerClass toolbox can be found at www.perclass.com.