magistrsko delo
Daniel Hari (Author), Matej Rojc (Mentor), Danilo Zimšek (Co-mentor)

Abstract

V magistrskem delu so predstavljeni pristopi ocenjevanja zaupanja v globokih nevronskih mrežah na primeru razpoznave števk. Ti pristopi nam omogočajo izboljšavo kakovosti razpoznave števk, s čimer se približamo natančnosti človeka, ki za bazo MNIST znaša 97,5–98 %. V delu se bomo osredotočili predvsem na dva pristopa, in sicer z Bayesovim učenjem in vzorčenjem z izpustnimi sloji. Bayesovo učenje je matematično bolj zahteven postopek, saj deluje tako, da vsak vhodni podatek v nevronsko mrežo obravnavamo kot porazdelitev verjetnosti in ne kot deterministično določeno vrednost. Pri tehniki vzorčenja z izpustnimi sloji je za vsakim skritim slojem mreže dodan stohastični izpustni sloj, tako da lahko na izhod iz modela gledamo kot na naključni vzorec, ki je ustvarjen iz aposteriorne porazdelitve. Takšen postopek je sicer računsko manj zahteven, daje pa podoben rezultat. Magistrsko delo je sestavljeno iz teoretičnega in eksperimentalnega dela. V teoretičnem delu so predstavljeni pojmi, kot so umetna inteligenca in sestava nevronske mreže ter podroben opis Bayesovega učenja in vzorčenja z izpustnimi sloji. V eksperimentalnem delu so prikazani pristopi razpoznave števk z Bayesovim učenjem in pristopi, ki uporabljajo tehnike vzorčenja z izpustnimi sloji. Podana je tudi primerjava postopkov.

Keywords

umetna inteligenca;Bayesov pristop;izpustni sloji;strojno učenje;magistrske naloge;

Data

Language: Slovenian
Year of publishing:
Typology: 2.09 - Master's Thesis
Organization: UM FERI - Faculty of Electrical Engineering and Computer Science
Publisher: [D. Hari]
UDC: 004.85(043.2)
COBISS: 49199107 Link will open in a new window
Views: 489
Downloads: 97
Average score: 0 (0 votes)
Metadata: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Other data

Secondary language: English
Secondary title: Confidence estimation in deep neural networks
Secondary abstract: The master's thesis presents approaches to assessing trust in deep neural networks in the case of digit recognition. These approaches allow us to improve the quality of digit recognition, thus approaching the accuracy of a human which ranges from 97.5% to 98% for the MNIST base. In the paper, we will focus mainly on two approaches, namely Bayesian learning and discharge layer sampling. Bayesian learning is a more mathematically demanding procedure because it works by treating each input data into a neural network as a probability distribution and not as a deterministically determined value. In the discharge layer sampling technique, a stochastic discharge layer is added behind each hidden layer of the grid. Thus, the output from the model can be viewed as a random sample created from the a posteriori distribution. Such a procedure is less computationally demanding but gives a similar result. The master's thesis consists of theoretical and experimental work. The theoretical part presents concepts, such as artificial intelligence and neural network composition, as well as a detailed description of Bayesian learning and discharge layer sampling. The experimental part presents digit recognition approaches with Bayesian learning and approaches that use discharge layer sampling techniques. A comparison of procedures is also given.
Secondary keywords: artificial intelligence;Bayesian method;dropout;machine learning.;
Type (COBISS): Master's thesis/paper
Thesis comment: Univ. v Mariboru, Fak. za elektrotehniko, računalništvo in informatiko, Telekomunikacije
Pages: XIII, 85 f.
ID: 12026992