Algoritmi strojnog učenja

master's thesis

• Author: Ema Dogančić
• Mentor: Zoran Škoda (mentor)
• Committee member: Zoran Škoda (predsjednik povjerenstva)
• Committee member: Igor Ciganović (član povjerenstva)
• Committee member: Ana Prlić (član povjerenstva)
• Committee member: Matija Bašić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Science; (Department of Mathematics); Zagreb
• Defense date and country: 2019-11-26, Croatia
• Scientific / art field, discipline and subdiscipline: NATURAL SCIENCES; Mathematics

Abstract

Ovaj diplomski rad bavi se klasičnim i konvolucijskim neuronskim mrežama. Cilj rada bio je opisati i objasniti način na koji se pokušava simulirati rad ljudskih neuronskih mreža na računalne sustave. Sve to s ciljem stvaranja programa koji mogu imitirati ljudsko razmišljanje i djelovanje. Iz tog razloga smo detaljno opisali klasične umjetne neuronske mreže s propagacijom unaprijed pri čemu smo posebnu pozornost posvetili backpropagation algoritmu. U zadnjem poglavlju prikazujemo konvolucijske neuronske mreže, njihovu građu pomoću konvolucijskih slojeva, slojeva sažimanja i potpuno povezanih slojeva te hiperparametara mreže. Dalje govorimo o prednostima konvolucijskih nad klasičnim neuronskim mrežama. Na samom kraju rada spominjemo neke od poznatijih konvolucijskih neuronskih mreža te govorimo o problemu prenaučenosti koji se često javlja u dubokim neuronskim mrežama.

Abstract (english)

This thesis addresses feed-forward and convolutional neural networks. The task of the thesis was to explain and describe the way human neural networks are simulated onto computer systems for the purpose of creating computer programs that are able to imitate human behavior. We provided a detailed description of feed-forward neural networks with emphasis on backpropagation algorithm. The last chapter is focused on presenting convolutional neural networks, their architecture consisting of convolutional layers, pooling layers, fully connected layers and hyperparameters in the network. Furthermore, we discuss advantages of convolutional neural networks over feed-forward neural networks. At the very end of the thesis we mention some of the more famous convolutional neural networks and we’re discussing the problem of overfitting, which often occurs in deep neural networks.

• Keywords: neuronske mreže; backpropagation algoritam; konvolucijske neuronske mreže; duboke neuronske mreže
• Keywords (english): neural networks; backpropagation algorithm; convolutional neural networks; deep neural networks
• Language: croatian
• URN:NBN: urn:nbn:hr:217:561211
• 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: 2020-07-01 12:46:49