M14 LEUVEN

In recent years, deep learning has advanced at a tremendous pace and is now the core methodology behind cutting-edge technologies such as image classification and captioning, autonomous driving, natural language processing and generation. One exciting and challenging application field for deep learning is the learning of dynamical systems, also known as system identification. In this field, tailor-made model architectures and fitting criteria should be designed to: retain structural physical knowledge when available; introduce regularization to avoid overfitting or enforce known relationships among variables; and optimize training efficiency by leveraging parallelization as much as possible. The objective of this course is to introduce deep learning concepts and recently developed tools for system identification, including meta learning concepts for automated learning of the identification procedure itself. The course combines theoretical lectures and hands-on practical sessions in PyTorch. Outline 1. Introduction to system identification and deep learning 2. Feedforward and recurrent neural networks for system identification 3. Numerical optimization algorithms for training neural networks 4. Integrating system theory in deep learning 5. Meta-learning Target audience Students, researchers, and practitioners who want to understand how complex system identification problems can be formulated and solved using modern deep learning techniques. Basic knowledge of Python can be beneficial to better follow hands-on practical sessions.