- This event has passed.
MICDE Seminar: Udo von Toussaint, PD, Group Leader at the Max-Planck-Institute for Plasmaphysics in Garching, Divison Numerical Methods for Plasmaphysics
March 18, 2021 @ 11:00 am - 12:00 pm
Bio: Dr. Udo v. Toussaint earned his PhD in Physics at the University of Bayreuth in 2000. He then worked as a Postdoctoral Researcher at NASA Ames (RIACS), in Mountain View, CA from 2000-2002. Since 2003, Dr. von Toussaint has been a Scientist at the Max-Planck Institute for Plasmaphysics in Garching. Dr. von Toussaint is also editor of the ‘Entropy’ journal.
Besides plasma-wall interaction, his research interests are focussed on the design of optimal analysis and measurement strategies (Bayesian experimental design) for computer- and physics experiments. This encompasses modern concepts of uncertainty quantification (UQ) of complex computer codes (e.g. Plasma-wall simulations) as well as active-learning systems, which dynamically decide which action (e.g. measurement of a specific spectral line) might yield the most informative data based on the results from previous actions. This is addressed with Machine Learning techniques, e.g. Hidden Markov Models (HMM), neutral networks or bayesian acyclic graphs and complemented by numerical methods like Markov Chain Monte Carlo (MCMC), sequential optimization or polynomial chaos expansion.
A BAYESIAN APPROACH TO ARTIFICIAL NEURAL NETWORKS: Artificial Neural networks (ANN) are famous for their advantageous flexibility for problems when there is insufficient knowledge to set up a proper model. On the other hand, this flexibility can cause overfitting and can hamper the generalization and stability of ANNs. Many approaches to regularize ANNs have been suggested (e.g. L1- or L2-norm based regularization) but most of them are based on ad hoc arguments. Employing the principle of transformation invariance, a general prior for feed-forward networks can be derived. This regularization prior not only favours cell and layer pruning but enable also a consistent Bayesian approach: Relying on Occam’s razor we demonstrate (as a proof of concept) how an ANN can be applied even in the >absence< of available training data. The relation to the concept of automatic relevance detection will be discussed.
The MICDE Winter 2021 Seminar Series is open to all. University of Michigan faculty and students interested in computational and data sciences are encouraged to attend.
Dr. von Toussaint will be hosted by Professor Xun Huan, Assistant Professor of Mechanical Engineering.
Register for this event via Zoom to receive an email with the link and passcode to connect. Note: You may register after the event has started.
Questions? Email MICDEemail@example.com