I.Title:Dependable Artificial Intelligence Systems
II.Presenter:Prof.PedroReviriego,Universidad Politécnica deMadrid
III.Time: 10:00-11:00, November 18th，2025
IV.Location:KC216
V.Host: Prof.LIUShanshan,SICE
VI.Overview
Artificial Intelligence (AI) systems are increasingly embedded in applications that demand high reliability, from autonomous vehicles and robotics to critical infrastructure and healthcare. However, the underlying electronic platforms that support these systems are inherently prone to faults caused by radiation, ageing, noise, manufacturing defects, and even malicious attacks. Such errors can compromise both system functionality and safety, making dependability a key design concern for next-generation AI systems. This talk will explore the fundamental challenges of building dependable AI, focusing on how errors arise in complex hardware and large-scale machine learning models. It will present methodologies for evaluating system reliability through fault injection and discuss recent advances in efficient error detection and correction techniques. Examples will include fault resilience in models such as BERT and Stable Diffusion, as well as protection strategies for FPGA-based accelerators and large language models. Finally, the session will highlight new opportunities for interdisciplinary research in dependable computing and offer practical insights into how to approach academic research effectively.
VII.About the presenter:
Pedro Reviriego is a Professor at the Universidad Politécnica de Madrid (UPM), where he leads research on dependable and efficient computing systems. His work focuses on reliability and fault tolerance in electronic and AI-based systems, including error detection and correction, hardware security, and resilient architectures. He has co-authored more than 200 papers in leading international journals and conferences and has collaborated with industry and academic partners worldwide. Pedro is a Senior Member of the IEEE and has served as an associate editor for IEEE and IET journals.
