Séminaire virtuel: vendredi 21 juin 2024
Misbah Razzaq (INRAe Tours)

Lien Zoom

  • Meeting ID: 867 6409 6440
  • passcode: 149120

13h00 - 13h50 – Misbah Razzaq (INRAe Tours)

Exploring the role of Answer Set Programming in explainable AI (XAI)

Following the notable success of artificial neural network-based approaches and their extensive applications in the biological domain, explainable AI has become a vibrant area of research. This surge in interest stems from the necessity to get insight into the underlying biological structure of data and to ground any clinical translation. Over the past decade, numerous machine learning methods, primarily heuristic-based, have been developed to tackle the problem of explainability. Due to the uncertainty of the proposed explanations as well as vulnerabilities of underlying models, explainability remains an open area of research. Recently, approaches based on symbolic AI-based techniques have been introduced. Due to the inherent interpretability and explicit representation of knowledge and reasoning processes, they allow for exhaustively enumerating explanations and identifying minimal ones. In this talk, we will present a symbolic approach based on answer set programming (ASP) for computing explanations. First, we will formulate the task of explanation, why it is necessary, and how to formulate a neural network in the form of a logic program. We will adapt a deletion-based algorithm to identify which combinations of features in the input data are most influential in determining the output of the underlying model. We will present the results of our approach on six different benchmark scenarios: heart disease diagnosis, thyroid recurrence, breast cancer, diabetes, E.\ coli promoter, and voting. Furthermore, we will compare our results with machine learning-based algorithm as well as existing logic-based techniques. Finally, we will highlight the possible future directions in this line of research.

Dernière modification le 21/06/2024