A Practical Introduction to Deep Probabilistic Modeling

François Lanusse (UMR AIM, CEA/CNRS/Université Paris-Saclay/Université de Paris)

In recent years Deep Learning has emerged as an extremely powerful tool with the potential of tackling many problems in the physical sciences, where the quality and volumes of data are becoming very challenging for classical techniques.

One of the most crucial aspects of using Deep Learning for science however, is to properly account for and model uncertainties, an aspect that is still often overlooked in the more general machine learning literature.

The aim of this session will be to provide an introduction to Deep Probabilistic Modeling, combining Deep Learning with Probabilities, and demonstrating that neural networks can be used and interpreted in a sound Bayesian context. In addition to presenting the mathematical background behind these claims, the session will be accompanied by a notebook introducing the software tools (JAX, TensorFlow Probability) needed to apply these methods in practice.

Link to slides: https://slides.com/eiffl/deep-probabilistic-learning-acd756