Context. The emerging field of neuro-symbolic artificial intelligence (AI) aims to integrate symbolic reasoning with neural networks to develop AI systems that exhibit more human-like reasoning capabilities [1-9]. By combining symbolic reasoning with connectionist learning, these approaches seek to leverage the learning and generalization strengths of neural networks while incorporating structured knowledge and logical rules for enhanced explainability and robustness.
The objective of this internship is to study and evaluate neuro-symbolic approaches based on neural probabilistic programming [1,2,4] in terms of performance, explainability, and applicability. The intern will conduct a comprehensive literature review, implement and test several methods (e.g., using DeepProbLog [1] and Scallop [2]), and analyse their strengths and limitations. A second part of the internship will be devoted to exploring how these approaches can be extended to handle complex ontology-based query answering.
1] Robin Manhaeve, Sebastijan Dumancíc, Angelika Kimmig,Thomas Demeester, and Luc De Raedt,Neural probabilisticlogic programming in DeepProbLog, Artificial Intelligence298(2021), 103504
[2] Ziyang Li, Jiani Huang, and Mayur Naik,Scallop: A language for neurosymbolic programming, Proc. ACM Program. Lang.7(2023), no. PLDI.
[3] scallop:https://www.scallop-lang.org/tutorial.html
[4] S Badreddine et al., Logic Tensor Networks, Artificial Intelligence (303), 2022
[5] R Riegel at al., Logical Tensor Networks, Artificial Intelligence (307), 2023
[6] Li et al., Learning with logical constraints but without shortcut satisfaction, Proc. ICLR 2023
[7] Xu et al., A neurosymbolic approach for encoding forst order logic constraints, Proc. ICLR 2024
[8] Hu et al., Harnessing Deep Neural Networks with Logic Rules, Proc ACL 2016
[9] Ahmed et al., Semantic Probabilistic Layers for Neuro-Symbolic Learning, Proc NIPS 2022
Tasks.
- Literature Review:
* Study key research works in the field of neuro-symbolic approaches.
* Identify the different paradigms and methods used (e.g., DeepProbLog, Logic Tensor Networks, ...).
- Identify use cases.
- Compare the performance of the models in terms of accuracy, training speed, and generalization ability.
- Evaluate the explainability of the models by analysing how decisions are made and justified.
- Analyse the results to identify the strengths and weaknesses of each approach.
- Study the extension to ontology-based query answering
This internship could lead to a PhD on this subject.
Contacts: (farouk.toumani / vincent.barra)@uca.fr
(c) GdR IASIS - CNRS - 2024.