1
Research Topic for the ParisTech/CSC PhD Program *Field (cf. List of fields below): Information and Communication Sciences and Technologies Subfield (Applied Physics, Chemistry, Mathematics, Mech. Eng. etc): Computer Engineering, Robotics, Artificial Intelligence. Title: Explainable neural-symbolic learning ParisTech School: ENSTA ParisTech Advisor(s) Name: Ass. Prof. Natalia Diaz Rodriguez Advisor(s) Email: [email protected] (Lab, website): asr.ensta-paristech.fr https://flowers.inria.fr/ Short description of possible research topics for a PhD: Current Deep Learning methods are powerful but data-hungry, sample inefficient and hard to interpret. On the other hand, classic knowledge based approaches such as logics, or rule based systems are endowed with less generalization capabilities. However, the latter are fully explainable. The objective of this project is to combine the best of both paradigms by using recent advances on relational deep learning and symbolic reinforcement learning to make black box neural models more explainable. For this purpose, specific datasets will be created in order to test abilities for a neural network to learn spatial and relational ontological properties that are inherent to description logics.  Data, tasks and methods will be proposed to ameliorate the problem of standard memory architectures struggling at tasks that heavily involve understanding the way in which entities are connected, estimating cause- effect relations, or conditional independences. The learning of finer grained semantic levels of description logics will be tested for explainability in computer vision real-life applications in order to produce interpretable rules or hypotheses to allow 1) more trusted AI models, and 2) models closer to human common-sense reasoning. Required background of the student (Which should be the main field of study of the applicant before applying): Computer Science/Engineering, Programming, Deep learning and Reinforcement Learning A list of 5(max.) representative publications of the group: (Related to the research topic) [1] T Lesort, N Díaz-Rodríguez, D Goudou, Jean-François, Filliat. S-RL Toolbox: Environments, Datasets and Evaluation Metrics for State Representation Learning. 2018. [2] Raffin, A Hill, R Traoré, T Lesort, N Díaz-Rodríguez, D Filliat State Representation Learning for Control: An Overview. 2018.  arXiv:1809.09369 [3 S Doncieux, D Filliat, N Diaz-Rodriguez, T Hospedales, R Duro, A Coninx, … Open-ended Learning: a Conceptual Framework based on Representational  Redescription. 2018. Frontiers in Neurorobotics 12, 59 [4] M Garnelo, K Arulkumaran, M Shanahan. Towards deep symbolic reinforcement learning Deep Reinforcement Learning . NIPS Workshop 2016. [5] Evans, R. and Grefenstette, E. Learning explanatory rules from noisy data. 2018.  Journal of Artificial Intelligence Research , 61:1–64.

Research Topic for the ParisTech/CSC PhD Program · Learning: a Conceptual Framework based on Representational Redescription. 2018. Frontiers in Neurorobotics 12, 59 [4] M Garnelo,

  • Upload
    others

  • View
    1

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Research Topic for the ParisTech/CSC PhD Program · Learning: a Conceptual Framework based on Representational Redescription. 2018. Frontiers in Neurorobotics 12, 59 [4] M Garnelo,

Research Topic for the ParisTech/CSC PhD Program

*Field (cf. List of fields below): Information and Communication Sciences and Technologies Subfield (Applied Physics, Chemistry, Mathematics, Mech. Eng. etc): Computer Engineering,Robotics, Artificial Intelligence.Title: Explainable neural-symbolic learningParisTech School: ENSTA ParisTechAdvisor(s) Name: Ass. Prof. Natalia Diaz RodriguezAdvisor(s) Email: [email protected](Lab, website): asr.ensta-paristech.fr https://flowers.inria.fr/

Short description of possible research topics for a PhD: Current Deep Learning methods are powerful but data­hungry, sample inefficient and hard to interpret.On the other hand, classic knowledge based approaches such as logics, or rule based systems areendowed with less generalization capabilities. However, the latter are fully explainable. The objectiveof this project is to combine the best of both paradigms by using recent advances on relational deeplearning and symbolic reinforcement learning to make black box neural models more explainable. Forthis purpose, specific datasets will be created in order to test abilities for a neural network to learnspatial and relational ontological properties that are inherent to description logics.   Data, tasks andmethods will be proposed to ameliorate the problem of standard memory architectures struggling attasks that heavily involve understanding the way in which entities are connected, estimating cause­effect   relations,   or   conditional   independences.   The   learning   of   finer   grained   semantic   levels   ofdescription logics will be tested for explainability in computer vision real­life applications in order toproduce interpretable rules or hypotheses to allow 1) more trusted AI models, and 2) models closer tohuman common­sense reasoning.

Required background of the student (Which should be the main field of study of the applicant before applying): Computer Science/Engineering, Programming, Deep learning and Reinforcement LearningA list of 5(max.) representative publications of the group: (Related to the research topic)[1] T Lesort, N Díaz­Rodríguez, D Goudou, Jean­François, Filliat. S­RL Toolbox: Environments, Datasets and Evaluation Metrics for State Representation Learning. 2018.[2] Raffin, A Hill, R Traoré, T Lesort, N Díaz­Rodríguez, D Filliat State Representation Learning for Control: An Overview. 2018.  arXiv:1809.09369[3 S Doncieux, D Filliat, N Diaz­Rodriguez, T Hospedales, R Duro, A Coninx, … Open­ended Learning: a Conceptual Framework based on Representational  Redescription. 2018. Frontiers in Neurorobotics 12, 59[4] M Garnelo, K Arulkumaran, M Shanahan. Towards deep symbolic reinforcement learning Deep Reinforcement Learning . NIPS Workshop 2016.[5] Evans, R. and Grefenstette, E. Learning explanatory rules from noisy data. 2018.  Journal of Artificial Intelligence Research , 61:1–64.