Interpretabilnost modela dubokog učenja

master's thesis

• Author: Domagoj Ravlić
• Mentor: Tomislav Šmuc (mentor)
• Committee member: Tomislav Šmuc (predsjednik povjerenstva)
• Committee member: Pavle Pandžić (član povjerenstva)
• Committee member: Filip Najman (član povjerenstva)
• Committee member: Mladen Vuković (član povjerenstva)
• Granter: University of Zagreb; Faculty of Science; (Department of Mathematics); Zagreb
• Defense date and country: 2019-12-04, Croatia
• Scientific / art field, discipline and subdiscipline: NATURAL SCIENCES; Mathematics

Abstract

Duboke neuronske mreže postižu zavidne performanse u područjima poput klasifikacije i predikcije. U današnje vrijeme čak i dostižu sposobnosti samih ljudskih eksperata. No, još uvijek se u najčešćem broju primjena ti modeli smatraju crnom kutijom za koje često ni sam autor ne može dovoljno dobro objasniti kako i zašto funkcioniraju. Upravo zbog toga se javlja potreba za tehnikama koje će na ljudski razumljiv način interpretirati modele, odnosno omogućiti ljudsku/ekspertsku procjenu njihove robusnosti. U ovom radu predstavljamo takve metode interpretabilnosti konvolucijskih neuronskih mreža što je još uvijek relativno novo područje istraživanja. Najviše proučavamo metode Grad-CAM i TCAV za koje provodimo i pripadne eksperimente. Osim njih spominjemo i metode vizualizacije CNN reprezentacije (ponajviše saliency mape), ”otpetljavanje” CNN reprezentacije u eksplanatorne grafove, učenje neuronske mreže koja ima interpretabilnu reprezentaciju te interpretabilnost middle-to-end učenja.

Abstract (english)

Deep neural networks achieve enviable performance in terms of classification and prediction. Nowadays we are even able to find those which are more efficient than human experts in certain fields. But, in a large number of applications, those models are considered as ”black box” since neither their author can’t explain them properly. Because of that, there is a need for techniques which can interpret models in a way understandable for humans and scientists as well. In this work, we present those methods of convolutional neural network interpretability which is still a young field of research. Most of all, we study Grad-CAM and TCAV methods for which we conduct experiments. Besides them, we study methods of CNN representation visualization (mostly saliency maps), ”disentangling” CNN representation into explanatory graphs, learning neural networks with interpretable representations and network interpretability for middle-to-end learning.

• Keywords: duboke neuronske mreže; metode interpretabilnosti
• Keywords (english): deep neural networks; interpretability methods
• Language: croatian
• URN:NBN: urn:nbn:hr:217:435149
• 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: 2020-07-14 12:12:17