Nowdays Information Technology is one of the most important part of our lives. The elemental reason for this is the accelerated, information-based society whose needs and tasks can be satisfied only by computer science. Our modern world has to face the fact, that there are more and more problems, that we need to solve, but just computers can give us right solution. The solution often lies in systems based on some empirical experience. So these system works as a model, that is why the result we get is informative for us.
Deep Neural Networks are such model systems as well. Their mechanism of action is inspired by the cells of the human brain (neurons), which are in close contact with each other, informations are transmitted between them, and they also process these informations. The same structure is characterized by Deep Neural Networks, whose perceptrons and layers also create, transmit, and process the incoming information.
The central topic of this paper is the learning ability of Deep Neural Networks, and the factors, that can infulance it. During my work I wanted to find out, how much reduce the efficiency of learning, if we manipulate a network of image sets with noises. I've done the manipulation in two different ways: with reduced number of training data set, and with a test data set transformed into grayscale. After that, I evaluated the impact of manipulation on learning efficiency.