Interpretiranje diskriminativnih konvolucijskih modela

master's thesis

• Title (english): Interpreting discriminative convolutional models
• Author: Tin Ceraj
• Mentor: Siniša Šegvić (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Slobodan Ribarić (član povjerenstva)
• Committee member: Vlado Glavinić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2020-07-09, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Interpretiranje dubokih konvolucijskih modela je bitno za njihovo korištenje u industriji kako ne bi došlo do neočekivanih tj. neželjenih ishoda. U ovom radu se proučavaju i testiraju metode za interpretaciju modela. Metode obuhvaćene ovim radom vizualiziraju modele generiranjem slika korišenjem znanja modela. Pokazano je da napredne metode generiraju realistične slike i da korišteni modeli ne uče samo jednostavne značajke, već da mogu razaznati klase kao cjeline koje i čovjek vidi. U eksperimentima je pokazano da su robusni modeli još bolji u tome, odnosno da oni razaznaju značajke na višoj razini od nerobusnih modela. Pokazani su rezultati eksperimenata i predloženi budući pravci razvoja.

Abstract (english)

Interpretation of deep convolutional models is essential for their use in industry in order to avoid unexpected, i.e. undesirable outcomes. This paper focuses on methods for interpreting models and test them. The methods covered in this paper are used to visualize a model by generating images using the model knowledge. It has been shown that advanced methods generate realistic images and that the models used not only learn simple features but that they can distinguish classes as wholes in a similar way a human can. Experiments have shown that robust models are even better at this, that they recognize features at a higher level than non-robust models. The results of the experiments are shown and future directions of development are suggested.

• Keywords: računalni vid; interpretacija modela; inverzija modela; generacija slika; optimizacija ulaza
• Keywords (english): computer vision; model interpretation; model inversion; image synthesis; input optimization
• Language: croatian
• URN:NBN: urn:nbn:hr:168:980572
• 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 09:50:36