Keywords: Deep learning, Rule extraction, Interpretability, EHR Internship Duration: 30/11/-1 - 30/11/-1
Head of the hosting team: Feng Chu
Website: Click here
Address of the host laboratory: IBISC Team AROB@S 23, Boulevard de France 91034 Evry France
Supervisor: Farida ZEHRAOUIE-mail: farida.zehraoui@univ-evry.fr Phone: +33164853464
Deep learning is a class of machine learning methods that can model data with different levels of abstraction. These methods are mainly based on artificial neural networks. They have enabled significant progress in several areas such as object recognition, signal analysis, automated natural language processing, etc. Despite their predictive power, deep neural networks are considered black boxes, which makes their interpretation difficult, especially in sensitive areas such as healthcare. Recently, there has been a growing interest in explaining these models [1], particularly in extracting rules for interpretation. Researchers have proposed various methods aimed at extracting symbolic decision rules from neural networks [2][3][4][5]. The project aims to study the existing methods and propose a new method for extracting rules from a deep neural network. The proposed approach will be applied to real data extracted from electronic records of patients admitted to intensive care units to predict “Sepsis” [6].
Project steps: • Study of different approaches to extracting rules from neural networks. • Propose and implement a new rule extraction method from a neural network that satisfies some desirable properties. • Applying the implemented method to real data from patients admitted to intensive care units to early prediction of “Sepsis”.
1.R.Guidotti,A.Monreale,S.Ruggieri,F.Turini,F.Giannotti,andD.Pedreschi.Asurveyof methods for explaining black box models. ACM Comput. Surv., 51(5):93:1–93:42, Aug. 2018. ISSN 0360-0300. 2.R.Andrews,J.Diederich,andA.B.Tickle.Surveyandcritiqueoftechniquesforextractingrules from trained artificial neural networks. Knowledge-Based Systems, 8(6):373 – 389, 1995. ISSN 0950-7051. Knowledge- based neural networks. 3.G.BolognaandY.Hayashi.Aruleextractionstudyonaneuralnetworktrainedbydeeplearning. In 2016 International Joint Conference on Neural Networks, IJCNN 2016, Vancouver, BC, Canada, July 24-29, 2016, pages 668–675. IEEE, 2016.
