Learning Compositional and Interpretable Neural Representations

Seminario Learning Compositional and Interpretable Neural Representations, tenuto dal dott. Riccardo Renzulli dell’Università di Torino.

Quando: venerdì 30 maggio, ore 9:10
Dove:
- in presenza: Aula 2I del Politecnico di Torino
- online al link: https://polito-it.zoom.us/j/86921497644?pwd=f4gIcUzeZYWXfWFfDzHsIAR6n3QLbn.1

Il seminario è svolto nell’ambito del corso Explainable and trustworthy AI, tenuto da Eliana Pastor, ed è correlato ai Corsi di Laurea Magistrale in “Ingegneria Informatica (Computer Engineering)” e “Data Science and Engineering” del nostro Dipartimento.

Abstract: While deep learning continues to advance state-of-the-art results, understanding how these models reason and represent information remains a key challenge. In this talk, we explore a spectrum of approaches to structuring and interpreting neural representations. The roadmap begins with logic-based reasoning and typicality in symbolic systems, then move to Capsule Networks, which aim to group neurons to preserve part-whole relationships in neural architectures. Finally, we turn to recent work on interpretability, covering concept bottleneck models and mechanistic interpretability, with a focus on methods for unlearning or erasing specific concepts from trained models. Throughout, we reflect on how structured neural representations can guide both performance and interpretability in AI.

Bio: Riccardo Renzulli is a postdoctoral researcher at the University of Turin, member of the EIDOS research group. His research focuses on artificial intelligence, with a particular interest in representation learning, interpretability, robustness, multimodal foundation models, and medical imaging. He earned his PhD in Computer Science from the University of Turin in 2023, focusing on capsule networks. He spent a visiting period at Aalto University, where he worked on visual localization for UAVs. Before the PhD, he worked as deep learning scientist in the industry.


Per ulteriori informazioni contattare:
Eliana Pastor: eliana.pastor@polito.it