Talk summary:

Recent theoretical and technical progress in artificial neural networks has significantly expanded the range of tasks that can be solved by machine intelligence. In particular, the advent of powerful parallel computing architectures, coupled with the availability of "big data'', allows to train large-scale, multi-layer neural networks known as deep learning systems. Further breakthroughs have been made possible by advances in neural network architectures, mostly thanks to the introduction of Transformers and diffusion models. These powerful systems achieve human-like (or even super-human) performance in challenging tasks that involve natural language understanding and image generation. In this seminar I will briefly review the foundations of modern deep learning systems, focusing in particular on generative AI models.

Lecturer bio:
Alberto Testolin received the M.Sc. degree in Computer Science (Artificial Intelligence) and the Ph.D. degree in Cognitive Science from the University of Padova, in 2011 and 2015, respectively. He has been Visiting Student and then Visiting Scholar at Stanford University. He is currently Assistant Professor at the University of Padova, focusing on cognitive modeling and AI. His main research interests include deep learning, generative models and neuro-symbolic systems, with the goal of building realistic models of visual perception, numerical cognition, and mathematical learning. Beside his primary interest in cognitive modeling, he also collaborates with computer scientists and electronic engineers to apply deep learning in signal processing and system optimization.