19 November 2024 – 4.30 pm
Online seminar

Third session of the Neuro-Symbolic Seminars (NeSS)

Speaker: Giuseppe Marra at the Catholic University of Leuven

The lack of transparency in the decision-making processes of deep learning systems poses a significant challenge for modern artificial intelligence (AI), as it undermines users’ ability to trust and verify these systems. To address this challenge, concept-based models (CBMs) have made significant progress by incorporating human-interpretable concepts into deep learning architectures. This approach allows predictions to be traced back to specific conceptual models that users can understand and act upon.

However, the task predictors of existing CBMs are not fully interpretable, which prevents in-depth analysis and any form of formal verification of their decision-making process prior to implementation, thereby raising significant reliability concerns. To address this shortcoming, we present the Concept Memory Reasoner (CMR), a new CBM designed to provide a human-understandable and verifiable task prediction process. Our approach consists of modelling each task prediction as a neural selection mechanism operating on a store of learnable logical rules, followed by a symbolic evaluation of the selected rule. The presence of an explicit store and the symbolic evaluation enable domain experts to inspect and formally verify the validity of certain global properties of interest to the task prediction process.

The experimental results show that DCR achieves a better balance between accuracy and interpretability than the most advanced CBMs, identifies logical rules consistent with basic truths, allows for the modification of rules, and enables verification prior to implementation.

Bio: Giuseppe Marra is an assistant professor in the Department of Computer Science at KU Leuven, where he is a member of the Declarative Languages and Artificial Intelligence (DTAI) research group. He was an FWO Post-Doc Fellow on the “Deep Statistical Relational Learning” project with Prof. Luc De Raedt. He obtained his PhD from the University of Florence in 2020 under the supervision of Prof. Marco Gori. His research interests focus on the integration of neural computation and relational reasoning, with a particular emphasis on logical and probabilistic reasoning. This work intersects with the fields of explainable AI, safe reinforcement learning, generative AI and geometric deep learning.

Contact person for the seminar: Rafael Peñaloza