Nadzirani pristupi za procjenu nesigurnosti predikcija dubokih modela

master's thesis

• Title (english): Supervised approaches for estimating uncertainty of deep model predictions
• Author: Ivan Grubišić
• Mentor: Siniša Šegvić (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Zoran Kalafatić (član povjerenstva)
• Committee member: Slobodan Ribarić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2018-07-18, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Ovaj rad razmatra problem procjene nesigurnosti predikcija kod dubokih modela. Obrađene su dvije skupine pristupa procjene nesigurnosti. Cilj jedne skupine pristupa je postići pouzdanu procjenu nesigurnosti i razlikovati je li uzrok nesigurnosti višeznačnost podataka ili nesigurnost u parametre aproskimacijom bayesovskih neuronskih mreža. Cilj druge skupine razmatranih pristupa je poboljšati procjenu pouzdanosti predikcije bez mijenjanja već naučene neuronske mreže i prepoznati je li ulazni primjer izvan razdiobe skupa za učenje ili je krivo klasificiran na temelju izlaza (razdiobe softmaksa ili logita). Opisani su i ispitani neki pristupi na zadacima semantičke segmentacije i klasifikacije slika.

Abstract (english)

This thesis considers the problem of predictive uncertainty estimation in deep models. Two classes of uncertainty estimation are cosidered. The goal of one of the classes is to achieve reliable uncertainty estiamtion and distinguish whether the cause of uncertainty is ambiguity of data or parameter uncertainty by approximating Bayesian neural networks. The goal of the other considered class of approaches is to improve predictive uncertainty estimation without changing a pre-trained neural network and to detect out-of-distribution or misclassified examples based on the output (softmax output or logits). Some approaches have been described and tested on semantic segmentation and image classification tasks.

• Keywords: duboko -učenje; strojno učenje; nesigurnost; statističko zaključivanje; varijacijsko zaključivanje; bayesovske neuronske mreže; prepoznavanje izvanrazdiobnih primjera; računalni vid
• Keywords (english): deep learning; machine learning; statistical inference; variational inference; Bayesian neural networks; out-of-distribution example detection; computer vision
• Language: croatian
• URN:NBN: urn:nbn:hr:168:311849
• 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
• Repository: FER Repository
• Created on: 2019-03-13 20:40:44