Comparison of Neural Networks Efficiency For Digit Recognition Trained By Gradient Methods and Evolutionary Algorithms

Umjetna neuronska mreža je skup povezanih samostalnih jedinica zvanih umjetnim neuronima. Osmišljena je po uzoru na ljudski mozak te nastoji imitirati njegove procese učenja. U ovom radu dodatno je pojašnjen pojam umjetne neuronske mreže te opisan postupak učenja na primjeru unaprijedne slojevite neuronske mreže. Taj tip mreže potom naučen je prepoznavati znamenke primjenom dvije različite metode optimizacije mreže. Odabrale su se metode optimizacije gradijentnim spustom te evolucijskim algoritmom. Metode su prvo opisane zatim korištene za učenje neuronske mreže. Izrađena je baza podataka s primjerima za učenje. Učenje dvjema metodama provedeno je nad istom bazom podataka. Navedene su metrike za ocjenjivanje rada umjetne neuronske mreže te prikazani i komentirani rezultati učenja.An artificial neural network is a set of connected stand-alone units called artificial neurons. It is modeled on the human brain and seeks to mimic its learning processes. This paper further clarifies the concept of an artificial neural network and describes its learning process on the example of a feedforward neural network. This type of network was then taught to recognize digits using two different optimization methods. The selected methods are gradient descent and evolutionary algorithm. The methods were first described and then used to optimize the neural network. A database with learning examples was created and used for both optimization methods. The main metrics for evaluating the performance of artificial neural networks are presented as well as the learning outcomes that are further commented on