Βαθιά νευρωνικά δίκτυα με μηχανές διανυσμάτων υποστήριξης στο επίπεδο εξόδου
In the present study, we examine the progress of the neural networks using the Support Vector Machines (SVM) in the output layer. SVM can provide accurate rates by determining the optimum decision boundary for data classification in two classes. The experiments were processed in two types of neural networks like Multilayer Perceptron (MLP) and convolutional neural network, using datasets MNIST, CIFAR-10 and CIFAR-100, within five hidden layers of numerous nodes. For the MNIST dataset, the accuracy that achieved based on two hidden layers were for the Multilayer Perceptron 0.9838 and for SVM 0.9848. For the CIFAR-10 dataset, the accuracy that achieved based on three hidden layers were for the MLP 0.454 and for SVM 0.49.Furthermore, for the CIFAR-100 dataset the accuracy that achieved based on two hidden layers were for the MLP 0.1857 and for SVM 0.2201. Finally, experiments were also conducted in the architecture of convolutional neural networks obtaining test errors (%) of 2.39% in the SVM and 2.46% in the convolutional network.