Oliver Paul

Tuesday, May 06, 2025, 11:00 - 11:45

TBA

Abstract:

In the past years, we have established Bayesian methods in the area of sensor calibration. The Bayesian approach applies to ensembles of sensor systems comprising a sensor element responding to a measurand of interest and additional sensor elements capturing information about parasitic influences to which it is cross-sensitive. The method has relied on a sufficiently rich model for the inference of the measurand of interest from the available sensor signals and an experimentally determined prior of the model parameters of the ensemble.

In this talk, I will report recent results on the following questions: Within the model space defined by the rich model, can one identify simpler models that possibly offer an even smaller predictive uncertainty than the rich model? The question is rooted in the intuition that smaller models can temper the uncertainty-contributing tendency of larger models to overfit experimental evidence. Historically, Ockham’s razor has recommended such model parsimony in more general terms.

Our answer is clear (with proof) [1]: Within the Bayesian framework, model order reduction always causes an information loss implying an increase of the predictive uncertainty. Nevertheless, the accuracy loss can be negligible for substantial model order reductions. Therefore, the situation is similar to that known from principal component analysis, where accuracy loss invariably parallels dimensional reduction.

We have applied the method to our previous demonstration case of Hall-stress-temperature sensor systems designed for measuring the magnetic field [1]. For calibration scenarios with 2, 4, and 6 calibration measurements, we identify 2-, 4-, and 5-parameter models with negligibly larger predictive uncertainty than the 11-paremeter model that has previously served for performing the sensor system’s inference task.

Finally, I will try to frame circumstances under which Ockham’s razor remains a valid principle.

[1] DOI: 10.1109/JSEN.2025.3549652

Short biography:

Oliver Paul received the Diploma degree in physics and the D.Sc. degree from ETH Zürich, Switzerland, in 1986 and 1990, respectively.

After postdoctoral work at the Fraunhofer Institute for Solar Energy Systems, Freiburg, Germany, he joined the Physical Electronics Laboratory, ETH Zürich, as a Lecturer and a Group Leader in 1992. Since 1998, he has been a Full Professor with the University of Freiburg, Germany, where he heads the Laboratory for Microsystem Materials, Department of Microsystems Engineering (IMTEK), Faculty of Engineering. Dr. Paul is a Co-Founder of Sensirion AG, Stäfa, Switzerland, and Atlas Neuroengineering, Leuven, Belgium. He was the Director of IMTEK and the Dean of the Faculty of Engineering, University of Freiburg, from 2006 to 2008 and from 2016 to 2018, respectively. He was a Founding Director of the German Cluster of Excellence BrainLinks-BrainTools, University of Freiburg. He holds various advisory positions at the University of Freiburg. He is a coauthor of more than 400 technical publications, patents, and books. The research of his group focuses on MEMS materials and fabrication technologies, physical microtransducers, and microstructures for industrial and life science applications.

Dr. Paul has been a member of the Editorial Board of Sensors and Actuators A: Physical and Journal of Micromechanics and Microengineering and the Editorial Advisory Board of the IEEJ Transactions on Electrical and Electronic Engineering. He cochaired the IEEE MEMS 2004 Conference.