27. Adhoc MAP uncertainties for normalizing flows


Peter Steinbach

Matter research at Helmholtz is a vast and heterogeneous academic field driven by experiment and simulation of unprecedented scale and quality. The Helmholtz AI consultant team led by Peter Steinbach supports these in extracting knowledge from image (radiograms or microscopy images) or table like datasets (Xray scattering, accelerator monitoring).

What is the data science project’s research question? When inverting simulations to obtain the posterior that produced observations, SGD can be used to obtain the maximum a posteriori prediction estimate, i.e. MAP estimate. As this represents the expected value, we wonder how do obtain the uncertainty or variance around this estimate. For this we would like to compare approaches using Stochastic Gradient Langevin Dynamics Optimization. We would like to study if this method yields results and compare against sampling methods.

What data will be worked on? Simulated data of accelerator monitoring and beam quality

What tasks will this project involve?  Implement a Stochastic Gradient Langevin Dynamics Optimization and compare to vanilla SGD

What is the expected outcome? Contribution to research paper

What infrastructure, programs and tools will be used? Can they be used remotely? The entire project can be done remotely using open-source software, pytorch and sbi in particular.

What skills are necessary for this project? Data analytics / statistics, Scientific computation, Computer simulations, Data mining / Machine learning, Deep learning

Is the data open source?  No

Interested candidates should be at Phd level. Peter Steinbach is looking for 1 visiting scientist, working on the project together with the team.