Detekcija izvandistribucijskih primjeraka suparničkim učenjem

master's thesis

• Title (english): Out-of-distribution detection by adversarial learning
• Author: Hrvoje Bušić
• Mentor: Siniša Šegvić (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Zoran Kalafatić (član povjerenstva)
• Committee member: Marko Čupić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2019-07-15, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Klasifikacija slika je važan zadatak računalnog vida koji u različitim primjenama postiže vrhunske rezultate, no korišteni diskriminativni konvolucijski modeli su skloni neopravdanom optimizmu što ih čini nespremnima za šire praktično prihvaćanje.
Razmotren je pristup učenju klasifikatora usmjeren ka razdvajanju prediktivnih distribucija unutar- i izvan-distribucijskih primjeraka za lakšu izgradnju detektora izvan-distribucijskih primjeraka zasnovanog na pragu, gdje se generativni suparnički model koristi za uzorkovanje najkorisnijih izvan-distribucijskih primjeraka u fazi učenja.
Performanse dobivenih klasifikatora su ispitane uporabom standardiziranog skupa metrika za problem detekcije izvan-distribucijskih primjeraka preko više različitih izvan-distribucijskih skupova različite težine.

Abstract (english)

Image classification is an important task of computer vision that achieves state-of-the-art results with various applications, but the discriminative convolutional models used are prone to unjustified optimism, making them unsuitable for wider practical adoption.
A novel training method is considered which yields classifiers that more effectively separate predictive distributions of in- and out-of-distribution samples, enabling the construction of better threshold-based out-of-distribution detectors.
The robust approach harnesses generative adversarial networks during training to sample the most useful out-of-distribution samples to present to the classifier.
Obtained classifiers were evaluated using a standard set of metrics for measuring out-of-distribution detection performance on several unseen datasets of different complexity.

• Keywords: računalni vid; detekcija izvan-distribucijskih primjeraka; nadzirano učenje; suparničko učenje
• Keywords (english): computer vision; out-of-distribution detection; supervised learning; adversarial learning
• Language: croatian
• URN:NBN: urn:nbn:hr:168:579798
• 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: 2020-01-07 17:10:04