Ekstrakcija pravil iz globokih nevronskih mrež

Matjašec, Urška (author), Sadikov, Aleksander (mentor), Jamnik, Mateja (co-mentor), Shams, Zohreh (co-mentor)

Despite their high accuracy, deep neural networks’ main disadvantage is their lack of transparency, interpretability and explainability. This has prevented them from being widely used in safety and security critical systems, for example, healthcare. There have been many attempts to interpret neural networks via rule extraction. However, the existing algorithms either do not take into account the entire structure of the network, are not applicable on deep neural networks, or are very time and memory expensive. This thesis introduces a scalable, decompositional rule extraction algorithm, which can extract simple, easy to understand IF-THEN rules from deep neural networks and can deal with the multi-class classification problems. The extracted rules approximate the network's behaviour and explain its output layer in terms of the input features. The algorithm was applied in the healthcare domain, where explainability is crucial. It was tested on a data set of breast cancer patients (METABRIC), and evaluated on two tasks of binary and multi-class classification problems. The algorithm's performance was compared to two baselines: pedagogical C5.0 and the decompositional algorithm DeepRED. As expected, the pedagogical baseline outperformed both decompositional algorithms in time and memory complexity, number of extracted rules and their average length. However, our algorithm provided more accurate rules with a higher level of fidelity than the pedagogical baseline. Our algorithm also outperformed the decompositional baseline DeepRED in all perspectives.

• Language: English
• Keywords: explainability of neural networks, deep feedforward neural networks, rule extraction, decision rules, decompositional algorithm
• Work type: Master's thesis/paper
• Organization: FMF - Faculty of Mathematics and Physics
• Year: 2020
• PID: 20.500.12556/RUL-118104
• Publication date in RUL: 20.08.2020

Secondary language

• Language: Slovenian
• Title: Rule extraction from deep neural networks

Abstract

Kljub visoki napovedni točnosti imajo globoke nevronske mreže veliko pomanjkljivost, ki jim preprečuje široko uporabo v varnostnih in varnostno kritičnih sistemih, na primer v zdravstvu. Ta ovira je nezmožnost pojasnjevanja njihovih odločitev. V namen interpretacije nevronskih mrež z ekstrakcijo pravil je bilo razvitih že kar nekaj algoritmov, vendar imajo vsi vsaj eno izmed sledečih pomanjkljivosti: ignorirajo skrite nivoje nevronske mreže in tako zanemarijo informacije iz notranjosti mreže, niso uporabni na globokih nevronskih mrežah, ali pa so časovno in prostorsko preveč zahtevni. To magistrsko delo predstavi razširljiv, dekompozicijski algoritem za ekstrakcijo preprostih in razumljivih če-potem pravil iz globokih nevronskih mrež, ki je sposoben reševanja problemov večciljne klasifikacije. Izvlečena pravila aproksimirajo obnašanje nevronske mreže in razlagajo izhodni nivo mreže z vhodnimi atributi. Algoritem smo uporabili na zdravstveni domeni, kjer je razložljivost ključna, in ga testirali na podatkovni bazi METABRIC, ki zajema podatke o pacientkah z rakom dojk. Algoritem smo uporabili za reševanje problemov binarne in večciljne klasifikacije in ga primerjali z dvema osnovnicama: pedagoškim algoritmom C5.0 in dekompozicijskim algoritmom DeepRED. Po pričakovanjih se je pedagoški algoritem C5.0 izkazal najbolje v časovni in prostorski zahtevnosti, številu izvlečenih pravil in njihovi povprečni dolžini. Vendar pa je naš algoritem zgradil pravila z boljšo napovedno točnostjo in večjo zvestobo nevronski mreži kot pedagoška osnovnica. Naš algoritem je v vseh pogledih tudi prekosil dekompozicijski algoritem DeepRED.

• Keywords: razložljivost nevronskih mrež, globoke usmerjene nevronske mreže, ekstrakcija pravil, odločitvena pravila, dekompozicijski algoritem