Fusion Science and Technology / Volume 62 / Number 2 / October 2012 / Pages 347-355
Selected Paper from the Seventh Fusion Data Validation Workshop 2012 (Part 1) / dx.doi.org/10.13182/FST12-A14626
This paper addresses the problem of finding a minimal and good enough training data set for classification purposes by using active learning and conformal predictors. Active learning means to have control in the selection process of training samples instead of choosing them in a random way. To this end, active learning methodologies look for establishing selection criteria in order to find out the samples that show better discrimination capabilities. In the present case, conformal predictors have been used for these purposes. Results will be presented in a five-class classification problem with images. The features are the vertical detail coefficients of the Haar wavelet transform at level four to diminish the sample dimensionality by reducing the spatial redundancy of the images. The active selection of training sets (through the reliability measures of a conformal predictor) allows the improvement of the classifiers. Here, the word "improvement" refers to obtaining higher generalization properties thereby avoiding overfitting. Support vector machines classifiers, in the one-versus-the-rest approach, have been used as the underlying classifiers.