Primjena dubokog učenja u obradi zvuka

master's thesis

• Author: Petra Vitez
• Mentor: Saša Singer (mentor)
• Committee member: Saša Singer (predsjednik povjerenstva)
• Committee member: Robert Manger (član povjerenstva)
• Committee member: Ilja Gogić (član povjerenstva)
• Committee member: Goran Muić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Science; (Department of Mathematics); Zagreb
• Defense date and country: 2018-09-26, Croatia
• Scientific / art field, discipline and subdiscipline: NATURAL SCIENCES; Mathematics

Abstract

U ovom radu objašnjavamo pojam neuronskih mreža i promatramo njihov matematički aspekt. Dajemo generalni uvid u arhitekturu neuronskih mreža, te način na koji neuronska mreža uči i kako se zatim ponaša na dosad neviđenim podacima. Zatim, dajemo naglasak na rekurentne neuronske mreže i kako one rješavaju problem klasifikacije nizova. Objašnjavamo posebnu arhitekturu rekurentnih neuronskih mreža, tzv. Long-Short Term Memory arhitekturu. Na kraju, pokazujemo kako smo iskoristili programski jezik Python i biblioteku otvorenog koda Tensorflow, za izradu modela rekurentne neuronske mreže, koji se sastoji od dvije Long-Short Term Memory jedinice, kako bismo riješili problem klasifikacije gradskih zvukova, na UrbanSound 8k skupu podataka. Koristimo metodu unakrsne validacije i, kao rezultate, prikazujemo prosječnu točnost koju dobivamo na testnim skupovima podataka, te matricu konfuzije za jedan testni skup podataka.

Abstract (english)

In this thesis we explain the concept of neural networks and observe their mathematical aspect. We give a general insight into the architecture of neural networks, the way the neural network learns, and how it behaves on previosly unseen data. Then, we give a special accent to recurrent neural networks and how they solve the problem of sequence classification. We explain a special architecture of recurrent neural networks, the so-called Long-Short Term Memory architecture. In the end, we present how we used Python programming language and the open source library Tensorflow, to create a recurrent neuronal network model that consists of two LongShort Term Memory units, to solve the problem of urban sound classification, on the Urban Sound 8k data set. We use the cross-validation method and, as results, we give the average accuracy obtained for test datasets, and the confusion matrix for one test dataset.

• Keywords: neuronska mreža; arhitektura neuronskih mreža; Python; Tensorflow; Long-Short Term Memory arhitektura
• Keywords (english): neural network; architecture of neural networks; Python; Tensorflow; Long-Short Term Memory architecture
• Language: croatian
• URN:NBN: urn:nbn:hr:217:474439
• 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: 2019-07-03 12:57:48