Fuzzification of training data class membership binary values for neural network algorithms

Abstract

We propose an algorithm improvement for classifying machine learning algorithms with the fuzzification of training data binary class membership values. This method can possibly be used to correct the training data output values during the training. The proposed modification can be used for algorithms running individual learners and also as an ensemble method for multiple learners for better performance. For this purpose, we define the single and the ensemble variants of the algorithm. Our experiment was done using convolutional neural network (CNN) classifiers for the base of our proposed method, however, these techniques might be used for other machine learning classifiers as well, which produce fuzzy output values. This fuzzification starts with using the original binary class membership values given in the dataset. During training these values are modified with the current knowledge of the machine learning algorithm.