Strukturirano duboko učenje pomoću neuronskih mreža temeljenih na grafovima

master's thesis

• Title (english): Structured Deep Learning with Graph Neural Networks
• Author: Toni Kukurin
• Mentor: Jan Šnajder (mentor)
• Committee member: Jan Šnajder (predsjednik povjerenstva)
• Committee member: Alan Jović (član povjerenstva)
• Committee member: Domagoj Jakobović (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2020-07-10, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Strukturirano razumijevanje svijeta jedna je od preostalih ljudskih prednosti nad današnjim algoritmima strojnog učenja. Percepcija svijeta u obliku povezanih dijelova do neke je mjere prirođena ljudskoj vrsti. Brojni zadaci računalnog vida i obrade prirodnog jezika trude se doseći takvo razumijevanje svijeta na temelju sirovih i minimalno označenih podataka. Ovaj se rad fokusira na strukturirano duboko učenje temeljeno na graf neuronskim mrežama. Rad opisuje postojeće pristupe učenju strukture sirovih podataka iz različitih modaliteta (slike i tekst) uz primjenu na zadatak pojačanog učenja uzemljenog u prirodnom jeziku. Kroz eksperimentalnu evaluaciju pokazujemo da graf neuronske mreže mogu biti koristan alat za učenje reprezentacija stanja u latentnom prostoru. Usprkos uspješnim rezultatima u domeni modeliranja svijeta, nalazimo da takve prikaze nije trivijalno prenijeti na zadatke učenja politike u pojačanom učenju bez modela.

Abstract (english)

Perceiving the world in a structured manner is one of the remaining advantages humans have over machine learning algorithms. Viewing the world as a sum of interconnected parts rather than an amorphous blob is something that comes naturally to us. A number of tasks in computer vision and natural language processing attempt to recover this structure based on raw data and minimally annotated structure. This thesis focuses on graph neural network based structured deep learning for natural language processing. We have studied the existing approaches to learning structure from raw data in different modalities (images and text) and applied some of the insights to a language-grounded policy learning task. Our experiments indicate graph neural networks can be a useful tool in learning object-segmented latent space transitions. Despite successes on the world modeling task, such pre-trained representations do not seem to trivially transfer to improved performance in downstream model-free policy learning algorithms.

• Keywords: strojno učenje; umjetna inteligencija; kontrastivno učenje; strukturirano modeliranje svijeta; graf neuronske mreže; obrada prirodnog jezika
• Keywords (english): machine learning; artificial intelligence; contrastive learning; structured world modeling; graph neural networks; natural language processing
• Language: english
• URN:NBN: urn:nbn:hr:168:547017
• 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: 2023-01-13 10:15:16