Duboki generativni modeli temeljeni na prostornom rasporedu dijelova slike

master's thesis

• Title (english): Deep generative models based on spatial layout of image parts
• Author: Matija Folnović
• Mentor: Siniša Šegvić (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Marin Golub (član povjerenstva)
• Committee member: Leonardo Jelenković (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2017-07-20, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Iako moderni duboki modeli raspoznaju slike bolje od ljudi, problem klasifikacije slika još ne možemo smatrati riješenim. Poznato je da su duboki modeli skloni precijeniti pouzdanost predikcije, što može dovesti do vrlo neugodnih pogrešaka. Taj problem može se umanjiti primjenom reprezentacija koje sadrže prostorni raspored dijelova slike. Konkretno, prenaučenost se može umanjiti korištenjem regularizatora koji prisiljava model da nauči rekonstruirati ulaznu sliku. Ova ideja je proučavana kroz objavljen model koji se zove Duboki Što-Gdje Autoenkoder (engl. Stacked What-Where Autoencoder). Istražujemo prednosti i mane ovog pristupa u kontekstu polunadzirane klasifikacije jednostavnih slika i prikazujemo eksperimentalne rezultate na skupu MNIST.

Abstract (english)

Although modern deep models classify images better than humans, image classification cannot be regarded as a solved problem yet. It's known that deep models tend to overestimate prediction accuracy, which can lead to unpleasent errors. This problem can be reduced by exploiting representations containing spatial layout of image parts. In particular, overfitting can be reduced by supplying a regularizer which forces the model to be able to reconstruct the input image. This idea has been studied in the previously published model known as Stacked What-Where Autoencoder. We research advantages and disadvantages of this approach in regards to semi-supervised classification of simple images and present experimental results on the MNIST dataset.

• Keywords: duboki generativni modeli; učenje reprezentacije; polunadzirano učenje
• Keywords (english): deep generative models; representation learning; semi-supervised learning
• Language: croatian
• URN:NBN: urn:nbn:hr:168:305338
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2019-03-13 20:15:42