Osnovni koncepti, metodologija i primjena objašnjive umjetne inteligencije

master's thesis

• Author: Matej Duvnjak
• Mentor: Tomislav Šmuc (mentor)
• Committee member: Tomislav Šmuc (predsjednik povjerenstva)
• Committee member: Zvonimir Bujanović (član povjerenstva)
• Committee member: Vedran Krčadinac (član povjerenstva)
• Committee member: Goran Igaly (član povjerenstva)
• Granter: University of Zagreb; Faculty of Science; (Department of Mathematics); Zagreb
• Defense date and country: 2020-09-29, Croatia
• Scientific / art field, discipline and subdiscipline: NATURAL SCIENCES; Mathematics

Abstract

U ovom radu predstavili smo dio metoda za Objašnjivost umjetne inteligencije, podijelili smo metode prema kategorijama. Objasnili terminologiju korištenu u ovom polju. Za većinu metoda uz teoretsko objašnjenje dali smo i primjere korištenja. Predstavili smo najvažnije Samo-objašnjive metode kao što su Linearna Regresija, Logistička regresija i Stabla odluke. Objasnili zašto ih smatramo samo-objašnjivim. Za Model-Agnostičke metode koje spadaju u Post-Hoc metode objašnjivosti naveli smo Graf djelomične ovisnosti, Grafove pojedinačnog uvjetnog očekivanja, Grafove Akumuliranih lokalnih učinaka, Lokalno interpretabilnih model Agnostičke (LIME), te SHAP vrijednosti. Sve ove metode, omogućavaju nam određeni uvid u rad bilo kojeg modela. Kod metoda koje su specifične za modele koncentrirali smo se na Konvolucijske Neuronske mreže, te prikazali poboljšanja u mapama vidljivosti. Također smo pokazali Objašnjive metode za CNN koje kao podatke imaju Vremenske nizove.

Abstract (english)

In this paper, we have presented some of the methods for Explanatory Artificial Intelligence, we have divided the methods into categories and explained the terminology used in this field. For most methods, in addition to a theoretical explanation, we have given examples of use. We presented the most important Self-Explainable methods that are Linear Regression, Logistic Regression and Decision Trees and explained why we consider them only explainable. For Model-Agnostic methods that belong to the Post-Hoc methods, we have explained the Graph of Partial Dependence, Graphs of Individual Conditional Expectation, Graphs of Accumulated Local Effects, Locally Interpretable Model Agnostic (LIME), and SHAP values. All these methods allow us to see any model. For model-specific methods, we concentrated on Convolutional Neural Networks, and showed which improvements exist in visibility maps of CNN. We also showed Interpretable methods for time-series CNN.

• Keywords: samo-objašnjive metode; Linearna Regresija; Logistička regresija; Stabla odluke
• Keywords (english): Explanatory Artificial Intelligence; Self-Explainable methods; Linear Regression; Logistic Regression; Decision Trees
• Language: croatian
• URN:NBN: urn:nbn:hr:217:942014
• Study programme: Title: Computer Science and Mathematics; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra računarstva i matematike (magistar/magistra računarstva i matematike)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access
• Created on: 2020-12-10 13:14:46