undergraduate thesis
Deep Learning with application in pattern classification

Bruno Bakula (2015)
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
Metadata
TitleDubinsko učenje s primjenom u klasifikaciji uzoraka
AuthorBruno Bakula
Mentor(s)Goran Martinović (thesis advisor)
Bruno Zorić (thesis advisor)
Abstract
U ovome radu objašnjen je koncept dubinskog učenja, kao naprednog postupka učenja različitih razina značajki podataka. Arhitektura dubinskih neuronskih mreža mnogo je složenija od arhitekture standardnih mreža, te sadrži minimalno dva skrivena sloja. Tri glavne arhitekture dubinskih mreža su naslagani autoenkoderi, deep belief networks i konvolucijske mreže. Naslagani autoenkoderi grade se od prethodno treniranih, običnih autoenkodera. Nakon postupka predtreniranja, cijela mreža se fino podešava čime je postupak treniranja završen. Traje višestruko duže u odnosu na treniranje standardnih mreže. Na provedenim eksperimentima ostvaren je značajan napredak u točnosti klasifikacije korištenjem naslaganih autoenkodera, u odnosu na standardne neuronske mreže. Potrebno je eksperimentirati sa različitim dimenzijama skrivenih slojeva, kako bi se pronašle optimalne postavke rada dubinskih mreža, za određeni skup podataka. Ključne riječi: autoenkoder, dubinsko učenje, naslagani autoenkoderi, neuronske mreže, pohlepno treniranje po slojevima
Keywordsautoencoder deep learning stacked autoencoders neural networks greedy layer-wise training
Parallel title (English)Deep Learning with application in pattern classification
Committee MembersGoran Martinović (committee member)
Hrvoje Glavaš (committee member)
Krešimir Nenadić (committee member)
GranterJosip Juraj Strossmayer University of Osijek
Faculty of Electrical Engineering, Computer Science and Information Technology Osijek
Lower level organizational unitsDepartment of Software Engineering
Chair of Programming Languages and Systems
PlaceOsijek
StateCroatia
Scientific field, discipline, subdisciplineTECHNICAL SCIENCES
Computing
Program Engineering
Study programme typeuniversity
Study levelundergraduate
Study programmeUndergraduate University Study Programme in Computer Engineering
Academic title abbreviationuniv.bacc.ing.comp.
Genreundergraduate thesis
Language Croatian
Defense date2015-07-09
Parallel abstract (English)
n this paper, the concept of deep learning as an advanced method for learning different levels of data features, was explained. Deep neural networks architecture is far more complex compared to standard network architectures, and it consists of minimum two hidden layers. Three main deep network architectures are stacked autoencoders, deep belief networks and convolutional networks. Stacked autoencoders are composed of normal autoencoders that were previously trained. After the pretraining step, the whole network is fine tuned, which finishes the training process. The training process duration is multiplied, regarding to standard networks training. A significant improvement in clasiffication accuracy was made in experiments using stacked autoencoders, in contrary to standard neural networks. Experimenting with different dimension of hidden layers is needed to find optimal settings for running deep networks, applied on specific data set.
Parallel keywords (Croatian)autoenkoder dubinsko učenje naslagani autoenkoderi neuronske mreže pohlepno treniranje po slojevima
Resource typetext
Access conditionOpen access
Terms of usehttp://rightsstatements.org/vocab/InC/1.0/
URN:NBNhttps://urn.nsk.hr/urn:nbn:hr:200:081141
CommitterAnka Ovničević