Master Thesis
Dissertation title: «Performance evaluation of neural network training algorithms and their applications» (in Greek).
Supervisor: Professor P. Pintelas, Department of Mathematics, University of Patras.
Dissertation title: «Performance evaluation of neural network training algorithms and their applications» (in Greek).
Supervisor: Professor P. Pintelas, Department of Mathematics, University of Patras.
Abstract: Artificial neural networks are parallel computational models comprised of densely interconnected, adaptive processing units, characterized by an inherent propensity for learning from experience and also discovering new knowledge. Due to their excellent capability of selflearning and selfadapting, they have been successfully applied in a wide spectrum of regions for the resolution of classification or regression problems such as biology, medicine, geology etc. In this work, we study the supervised training process of an artificial neural network which is considered
eminently suitable for case where the education allocated important time and requires big stocking space, as it often happens when we have big totals of models and/or networks. In the literature, there have been proposed several algorithms for neural network training, covering the one the voids of the other, drawn so that they solve the problems that older(s) was difficult to solve. The aim of this work is to evaluate the most popular training algorithms measuring their faculty of training and their generalization capability in a variety of benchmarks from medicine and bioinformatics. Additionally, we propose two training algorithms which preserve the theoretical advantages of the classical algorithms and simultaneously provide faster, more stable and reliable convergence. Moreover, motivated from the possibility for the achievement of better generalization results we investigate the contribution of artificial neural networks in machine learning. Concretely, we evaluate the contribution of neural network ensembles in the formation of reliable decision making systems. Finally, we study possibilities of their combination with various classifiers for developing more powerful hybrid systems for exporting
information.
eminently suitable for case where the education allocated important time and requires big stocking space, as it often happens when we have big totals of models and/or networks. In the literature, there have been proposed several algorithms for neural network training, covering the one the voids of the other, drawn so that they solve the problems that older(s) was difficult to solve. The aim of this work is to evaluate the most popular training algorithms measuring their faculty of training and their generalization capability in a variety of benchmarks from medicine and bioinformatics. Additionally, we propose two training algorithms which preserve the theoretical advantages of the classical algorithms and simultaneously provide faster, more stable and reliable convergence. Moreover, motivated from the possibility for the achievement of better generalization results we investigate the contribution of artificial neural networks in machine learning. Concretely, we evaluate the contribution of neural network ensembles in the formation of reliable decision making systems. Finally, we study possibilities of their combination with various classifiers for developing more powerful hybrid systems for exporting
information.