Metodi per l'apprendimento di ensemble di reti neurali profonde.

In this thesis, we study methods to create ensembles of deep neural networks. Ensembles of deep neural networks are sets of deep networks whose outputs are fused together to obtain a single result. The idea behind ensembles is that the overall prediction of multiple networks will be better than the ones of the single networks. We introduce many methods used in the literature to create ensembles and we divide them into different categories depending on how the classifiers are forced to learn different features of the input data. Our experiments show that using deep networks that are different from each other in their architectures, in their training or in their image preprocessing allows to reach better performances than those that are reached by baseline ensembles of well-performing but similar networks.