Varijacijski autoenkoderi s kvantiziranom latentnom reprezentacijom

undergraduate thesis

• Title (english): Variational autoencodets with quantized latent representation
• Author: Dominik Babić
• Mentor: Siniša Šegvić (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Ivan Grubišić (član povjerenstva)
• Committee member: Josip Šarić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2023-07-06, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Zadatak ovog rada je bio uvod u osnovne pojmove i koncepte VQ-VAE modela
te uspostaviti vezu s već definiranim varijacijskim autoenkoderima. Uz to, cilj je
bio predstaviti osnovne matematičke izvode za definiranje VQ-VAE modela, kao što
su varijacijska inferencija, rekonstrukcijski gubitak, regularizacijski gubitak i gubitak
kvantizacije.
Zatim je provedeno izvod̄enje varijacijskih autoenkodera i VQ-VAE modele te
njihovo treniranje na jednostavnijem skupu slika (konkretno, MNIST slika znamenki)
kako bi se istražila njihova sposobnost rekonstrukcije slika. Takod̄er je bilo potrebno
odabrati odred̄enu javno dostupnu prednaučenu arhitekturu i provesti ispitivanje nad
njom.
Potom su prikazani dobiveni rezultati provedenih ispitivanja te su izvedeni zaključci
o performansama ispitanih arhitektura.

Abstract (english)

This study aimed to provide an introduction to the fundamental concepts
and principles of the VQ-VAE model and establish a connection with the already
defined variational autoencoders. Additionally, the goal was to present the basic mathematical
derivations for defining the VQ-VAE model, such as variational inference, reconstruction
loss, regularization loss, and quantization loss.
Furthermore, variational autoencoders and VQ-VAE models were derived and trained
on a more specific dataset of images (specifically, MNIST digit images) to explore their
ability to reconstruct images. It was also necessary to select a specific publicly available
pre-trained architecture and evaluate it.
Subsequently, the obtained results of the conducted evaluations were presented,
and conclusions were drawn regarding the performance of the examined architectures.

• Keywords: VQ-VAE; varijacijski autoenkoderi; generativni modeli; varijacijska inferencija; rekonstrukcijski gubitak; kvantizacija
• Keywords (english): VQ-VAE; variational autoencoders; generative models; variational inference; reconstruction loss; quantization
• Language: croatian
• URN:NBN: urn:nbn:hr:168:456628
• Study programme: Title: Computing; Study programme type: university; Study level: undergraduate; Academic / professional title: sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva (sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-07-13 09:53:13