Speaker: Dr. Lars Heyen (KIT)
Title of Talk: Neural Networks with Error Bars – an Introduction to Bayesian Neural Networks
Abstract: Neural networks are used for an ever-increasing number of tasks in both science and everyday life. However, most of these networks do not provide any measure of their predictions’ certainty, which can lead to undetected hallucinations.
In this talk, I will give an introduction to Bayesian neural networks which integrate uncertainty quantification directly into the network structure and training procedure. Variational inference allows for a practical implementation of this concept, striking a good balance between the computational cost and the quality of the probabilistic prediction. I will demonstrate this technique on the task of weather forecasting and discuss some of the intricacies of Bayesian training.
Venue: Physikalisches Institut, Goldene Box
Schedule:
Talk: 16:30 – 17:15
Questions & Discussion: 17:15 – 17:30
Pizza & Drinks: 17:30
_____________________________________________________________________________________________