Duboke neuronske mreže

Duboke neuronske mreže su sve popularnije zahvaljujući velikim količinama podataka, većoj snazi računala te zajednici okruženoj oko softverskih alata. Koriste se za obradu računalnog vida, prirodnog jezika, tekstova te raznih predikcija. U ovom radu najprije opisujemo općeniti pojam neuronske mreže te neophodne segmente za učenje modela, a to su: podaci, mjere uspješnosti, arhitektura modela i algoritmi učenja. Dalje se uvode i analiziraju optimizacijski algoritmi, mrežne arhi- tekture i njihove primjene. Na kraju se pokazuje važnost dubine neuronskih mreža, rješenjem problema klasi.Deep neural networks are becoming popular because of a large amount of existing data, increased computing power as well as the community created around the software. Networks are used for computer vision, natural language processing and various predictions. This thesis, firstly describes the general concept of a neural network and the necessary segments for model learning, of which there are four: data, performance measures, model architecture and learning algorithms. Optimization algorithms, network architectures and their applications are further introduced and analyzed. Finally, the importance of the depth is shown by solving the problem of classifying images of handwritten characters