prikaz prve stranice dokumenta Automatski dizajn konvolucijske neuronske mreže uporabom optimizacije rojem čestica
  Download
PDF 1.38 MB
master's thesis
Automatski dizajn konvolucijske neuronske mreže uporabom optimizacije rojem čestica
Osijek: Josip Juraj Strossmayer University of Osijek, Faculty of Electrical Engineering, Computer Science and Information Technology Osijek, 2024. urn:nbn:hr:200:254626
Ukić, Ante
Josip Juraj Strossmayer University of Osijek
Faculty of Electrical Engineering, Computer Science and Information Technology Osijek
Department of Software Engineering
Chair of Programming Languages and Systems

Institutional repository: Faculty of Electrical Engineering, Computer Science and Information Technology Osijek

Cite this document

Ukić, A. (2024). Automatski dizajn konvolucijske neuronske mreže uporabom optimizacije rojem čestica (Master's thesis). Osijek: Josip Juraj Strossmayer University of Osijek, Faculty of Electrical Engineering, Computer Science and Information Technology Osijek. Retrieved from https://urn.nsk.hr/urn:nbn:hr:200:254626

Ukić, Ante. "Automatski dizajn konvolucijske neuronske mreže uporabom optimizacije rojem čestica." Master's thesis, Josip Juraj Strossmayer University of Osijek, Faculty of Electrical Engineering, Computer Science and Information Technology Osijek, 2024. https://urn.nsk.hr/urn:nbn:hr:200:254626

Ukić, Ante. "Automatski dizajn konvolucijske neuronske mreže uporabom optimizacije rojem čestica." Master's thesis, Josip Juraj Strossmayer University of Osijek, Faculty of Electrical Engineering, Computer Science and Information Technology Osijek, 2024. https://urn.nsk.hr/urn:nbn:hr:200:254626

Ukić, A. (2024). 'Automatski dizajn konvolucijske neuronske mreže uporabom optimizacije rojem čestica', Master's thesis, Josip Juraj Strossmayer University of Osijek, Faculty of Electrical Engineering, Computer Science and Information Technology Osijek, accessed 27 November 2024, https://urn.nsk.hr/urn:nbn:hr:200:254626

Ukić A. Automatski dizajn konvolucijske neuronske mreže uporabom optimizacije rojem čestica [Master's thesis]. Osijek: Josip Juraj Strossmayer University of Osijek, Faculty of Electrical Engineering, Computer Science and Information Technology Osijek; 2024 [cited 2024 November 27] Available at: https://urn.nsk.hr/urn:nbn:hr:200:254626

A. Ukić, "Automatski dizajn konvolucijske neuronske mreže uporabom optimizacije rojem čestica", Master's thesis, Josip Juraj Strossmayer University of Osijek, Faculty of Electrical Engineering, Computer Science and Information Technology Osijek, Osijek, 2024. Available at: https://urn.nsk.hr/urn:nbn:hr:200:254626

Metadata
TitleAutomatski dizajn konvolucijske neuronske mreže uporabom optimizacije rojem čestica
Title (english)Automatic design of convolutional neural network using particle swarm optimization
AuthorAnte Ukić
MentorDražen Bajer (mentor)
MentorMario Dudjak (sumentor)
Committee memberDražen Bajer (predsjednik povjerenstva)
Committee memberKrešimir Romić (član povjerenstva)
Committee memberBruno Zorić (član povjerenstva)
GranterJosip Juraj Strossmayer University of Osijek
Faculty of Electrical Engineering, Computer Science and Information Technology Osijek
(Department of Software Engineering)
(Chair of Programming Languages and Systems)
Osijek
Defense date and country2024-02-15, Croatia
Scientific / art field,
discipline and subdiscipline		TECHNICAL SCIENCES
Computing
Artificial Intelligence
Abstract
Konvolucijske neuronske mreže primjenjuju se za rješavanje raznih problema dubokog učenja. Primjena konvolucijskih neuronskih mreža za zadani problem zahtijeva dizajniranje odgovarajuće strukture mreže. Ručni odabir i vrednovanje potencijalnih struktura mreža zamoran je i dugotrajan proces. Automatizacija ovog postupka uvelike može smanjiti napore i vrijeme potrebno za dizajniranje prikladnih mreža. Automatski dizajn neuronskih mreža može se predstaviti kao optimizacijski problem te je u literaturi predloženo pregršt postupaka za njegovo rješavanje. U ovom radu korišten je algoritam PSO koji se pokazao učinkovitim u rješavanju brojnih problema optimizacije u literaturi, među kojima je i dizajn neuronskih mreža. Kako bi se performanse algoritma PSO mogle usporediti s nekom od poznatijih i jednostavnijih algoritama za dizajniranje neuronskih mreža, proveden je postupak pronalaska prikladne inačice algoritma PSO u smislu načina predstavljanja rješenja i odabira vrijednosti njegovih parametara. Eksperimentalna analiza provedena je na četiri skupa podataka: MNIST, Fashion MNIST, CIFAR-10 i CIFAR-100. Eksperimentalnom analizom utvrđeno je kako je najbolji način predstavljanja rješenja pomoću vektora realnih vrijednosti, a kod odabira parametara povoljnim se pokazalo odabrati jednake vrijednosti akceleracijskih koeficijenata. Eksperimentalnom analizom također je utvrđeno kako algoritam PSO postiže bolje rezultate od uobičajeno korištenih postupaka za dizajniranje neuronskih mreža, što ga čini prikladnim algoritmom za ovaj problem.
Abstract (english)
Convolutional neural networks are applied to solve various deep learning problems. Applying convolutional neural networks to a given problem requires designing an appropriate network structure. Manual selection and evaluation of potential network structures is a tedious and time-consuming process. Automating this process can significantly reduce the effort and time needed to design suitable networks. The automatic design of neural networks can be presented as an optimization problem, and a handful of procedures for solving it have been proposed in the literature. In this paper, the PSO algorithm was used, which proved to be effective in solving numerous optimization problems in the literature, among which is the design of neural networks. To compare the performance of the PSO algorithm with more well-known and simpler algorithms for neural network design, a process of finding a suitable version of the PSO algorithm in terms of solution representation and parameter value selection was conducted. Experimental analysis was carried out on four datasets: MNIST, Fashion MNIST, CIFAR-10 and CIFAR-100. Through experimental analysis, it was determined that the best way to represent solutions is through a vector of real values and selecting equal values for acceleration coefficients proved favorable in parameter selection. The experimental analysis also revealed that the PSO algorithm achieves better results than commonly used procedures for neural network design, making it a suitable algorithm for this problem.
Keywords
Keywords (english)
Languagecroatian
URN:NBNurn:nbn:hr:200:254626
Study programmeTitle: Graduate University Study Programme in Computer Engineering
Study programme type: university
Study level: graduate
Academic / professional title: sveučilišni/a magistar/magistra inženjer/inženjerka računarstva (sveučilišni/a magistar/magistra inženjer/inženjerka računarstva)
Type of resourceText
File originBorn digital
Access conditionsOpen access
Terms of useLicence In copyright icon
Created on2024-03-07 11:07:34