Visualizing learning of neural networks

V diplomski nalogi se osredotočamo na področje vizualizacije modelov umetnih nevronskih mrež. Trenutno še ni veliko razvitih rešitev, ki bi širši javnosti razlagale osnove delovanje modelov nevronskih mrež. Obstoječe rešitve so strokovnejše narave, ki pa so laikom težko razumljive. Zato smo z različnimi pristopi razvili vizualizacije poteka učenja in vpliva parametrov modela. Ustvarili smo interaktivno spletno stran, ki v ozadju zažene ustvarjen model večnivojskega perceptrona, nato pa s štirimi vizualizacijami prikazuje dogajanje znotraj modela med učenjem. Pri testiranju se je izkazalo, da spletna stran podaja anketirancem nov vpogled v delovanje nevronskih mrež, vendar pa določenih parametrov in funkcionalnosti še vedno ne pojasni dovolj.This diploma thesis focuses on the field of visualization of artificial neural network models. Currently, there is a lack of developed solutions for the general public, which explain the intrinsics of these models. Existing models are intended for scientific use, which makes it harder to understand for laymans. That is why we decided to use different approaches to develop visualizations of training the model and the influence of its parameters. We developed an interactive website, which runs a multilayer perceptron model in the background and displays four visualizations, with which we show what is happening in our model during training. Our testing results of the website indicate that it does deliver new aspects to understanding of neural network models, but still does not fully explain certain parameters and its functionalities