Artificial Intelligence Seminar - Dr. Hamed Ayoobi, Imperial College London

Title: Generative AI That Argues: Explanations for Neural Networks through Arguments, Prototypes, and Conflicts

Abstract:

Generative AI relies on diverse neural architectures—from MLPs and CNNs to modern Transformers—yet their reasoning remains hidden. In this talk, I will present a pathway towards contestable generative AI: systems that not only explain their outputs, but also argue for them and expose internal conflicts. I will introduce three recent contributions: (1) SpArX, which produces sparse argumentative explanations faithful to the inner logic of multilayer perceptrons; (2) ProtoArgNet, which grounds image classifications in super-prototypes and generates pro/con visual arguments for convolutional networks; and (3) our work on Hidden Conflicts, which detects contradictory evidence inside neural networks, including Transformers, that destabilize explanations. Together, these methods move beyond static saliency maps to enable interactive, revisable, and trustworthy explanations, with applications in fairness auditing and clinical decision support.

This seminar will be followed by a lecture of Dr. Hamed Ayoobi until 11:45 hours.