Thesis title: A Generalized learning approach to Deep Neural Networks
This thesis in the first part proposes a Machine Learning (ML) approach for the analysis and classi cation of Ground Penetrating Radar (GPR) given a limited number of B-scan images obtain with a dedicated simulation tool. Speci cally, both a custom Convolutional Neural Network (CNN) and a a well-established Deep Learning (DL) architecture, DenseNet, that is opportunely scaled-down to take into account the small dataset, are considered. Those networks are then
employed to classify B-scan simulations from buried cylinders in order to retrieve the host media permittivity, the cylinder depth respect to surface, and cylinders radius. The main aim of the proposed work is to test the applicability of a scaled down version of DenseNet architecture to the analysis of B-scan images and compare the performance respect to a classical CNN. The architecture chosen has shown interesting results in retrieving information from a limited images data set. The second part of the thesis presents a Generalized Newton’s Method as a powerful approach to learning in Deep Neural Networks.
This technique was compared to two popular approaches, namely the Stochastic Gradient Descent and the Adam algorithm, in two popular classification tasks. The performance of the proposed approach confirmed it as an attractive alternative to state-of-the-art first order solutions