Mots-Clés

Explainable AI, Answer set programming, Neural networks, logic programming

Description

Start and end dates are flexible

Exploring Hybrid Explainable AI Frameworks: Artificial Neural Networks and Answer Set Programming in Biological Applications

Explainable AI has become a vibrant area of research due to the need to understand the internal workings of artificial neural networks and to identify the most essential factors. This is a crucial precondition for gaining insight into the underlying biological structure of data and grounding any clinical translation. In the past decade, heuristic-based machine learning methods have emerged to address explainability, but uncertainty persists, keeping it an active research area. Symbolic AI approaches, with their inherent interpretability, now offer exhaustive and minimal explanations.

Recently, we adapted one such approach and use answer set programming (ASP) for computing explanations. We develop a logic program to formulate artificial neural networks and adapt a deletion based algorithm to identify which combinations of features in the input data are most influential in determining a network's output. Our results indicate that the ASP-based approach is competitive with other logic-based techniques and in some instances even generates smaller explanations, further emphasizing its potential in practical applications and highlighting the potential of combing neural networks with ASP to improve interpretability. 

The objective of this internship includes:

• Comprehensive Literature Review: Survey the existing logic based approaches for explaining neural networks and benchmarking their performance.
• Benchmarking and Comparative Analysis: Apply the ASP-based approach developed in the team to the larger case studies and its comparison with other approaches.
• Tool adaptation and development: Continuing the tool development to propose solutions for scalability and robustness issues.

Required skills:

• Technical Skills: Knowledge of Artificial Neural Networks, symbolic AI (particularly ASP), and machine learning.
• Programming: Proficiency in Python (mandatory) and logic programming (preferred).
• Tools: Familiarity with machine learning frameworks (mandatory) and ASP solvers (preferred).
• Domain Knowledge: A basic understanding of biological datasets and their challenges is preferred, but not mandatory.

Long-term goal: This internship offers an exciting opportunity to develop a hybrid approach that combines the predictive strength of artificial neural networks with the interpretability of Answer Set Programming (ASP). It aims to bridge the gap between computational techniques and their real-world applications in biology by advancing research in explainable AI.

Contact: Send an email to misbah.razzaq@inrae.fr including a motivation letter, academic grades, and the contact information of your previous supervisor.