An experimental approach founded on theoretical background is what allows us to find logical answers to fundamental questions, and gives us some important guidelines on how to choose a reasonable solution to the problem that we want to solve. By manipulating different parts of a neural network algorithm, it is possible to observe how its behaviour changes, and if this process is done by modifying one piece at a time it is possible to identify what are the reasons for this change, and judge why it is so. In this dissertation we aim to follow an experimental approach, and to understand and describe the design choices for solving a classification problem, focusing particularly on confronting different loss functions.
Go to the Dissertation