Konvolucijske neuronske mreže

master's thesis

• Author: Kristina Džomba
• Mentor: Saša Singer (mentor)
• Committee member: Saša Singer (predsjednik povjerenstva)
• Committee member: Mladen Jurak (član povjerenstva)
• Committee member: Vedran Čačić (član povjerenstva)
• Committee member: Maja Starčević (član povjerenstva)
• Granter: University of Zagreb; Faculty of Science; (Department of Mathematics); Zagreb
• Defense date and country: 2018-07-18, Croatia
• Scientific / art field, discipline and subdiscipline: NATURAL SCIENCES; Mathematics

Abstract

U ovom diplomskom radu detaljno je obrađen model konvolucijskih neuronskih mreža. U prvom dijelu rada dan je uvod u klasične neuronske mreže, radi boljeg razumijevanja arhitekture i osnovnih pojmova dubokih modela. Drugi dio rada započinje opisom operacije konvolucije, koja se primjenjuje u konvolucijskim mrežama. U nastavku smo opisali svaki sloj konvolucijske mreže te dali matematički prikaz algoritma učenja modela. Nakon toga, naveli smo neke od najčešćih problema koji se javljaju prilikom učenja mreže te prikazali neke od optimizacijskih i regularizacijskih metoda, koje rješavaju dane probleme i poboljšavaju učenje modela. Na samom kraju rada, opisali smo implementacijsko rješenje za problem neuronskog prijenosa stila, s umjetničke slike na odabranu fotografiju. Model je razvijen pomoću biblioteke TensorFlow za strojno učenje. Za učenje modela korištena je okolina razvojnog oblaka Google Colaboratory, koja pruža besplatnu upotrebu grafičkih kartica. Rezultati modela su slike s primijenjenim stilom drugih slika, koje su prikazane u posljednjem poglavlju rada.

Abstract (english)

In this thesis we gave a detailed view of convolutional neural network models. In the first part of the thesis, we presented a gentle introduction to classic neural networks for better understanding, with some basic concepts of deep learning models. Second part of the thesis begins with the convolution operation description, which is the main carrier operation in convolutional neural networks. Furthermore, we described every CNN layer, and gave a mathematical background for the learning algorithm. After that, we presented some most common problems which can occur while training deep models. We described some optimization and regularization methods that can solve these problems and improve the model learning. Lastly, we described an implementation for neural style transfer problem, from an artistic image to a chosen photography. The model is implemented using TensorFlow machine learning library. Model is learned in Google Colaboratory environment which offers free GPU support. Results of the model computations are images, with style transferred from other images, as shown in the last chapter of the thesis.

• Keywords: konvolucijske neuronske mreže; konvolucija; algoritam učenja modela; neuronski prijenos stila; TensorFlow; Google Colaboratory
• Keywords (english): convolutional neural network; convolution; learning algorithm; neural style transfer problem; TensorFlow; Google Colaboratory
• Language: croatian
• URN:NBN: urn:nbn:hr:217:020017
• Study programme: Title: Computer Science and Mathematics; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra računarstva i matematike (magistar/magistra računarstva i matematike)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access
• Created on: 2018-11-23 11:30:36