AI becomes even smarter through purposeful integration of specific prior knowledge

New Operator-Based Learning research group emerges from Fraunhofer Attract program

 

A new Operator-Based Learning group has been established at the Fraunhofer Development Center X-ray Technology EZRT at Fraunhofer IIS. The research group emerged from the Fraunhofer Attract program, which offers external applicants the opportunity to put forward their own ideas to a Fraunhofer Institute as research topics.

AI in image reconstruction

Thanks to machine learning, we can manage many tasks a good deal more efficiently and purposefully than before. The self-learning systems draw on a huge database of existing knowledge for their task and “condition” themselves to come up with ever better results.

This artificial intelligence (AI) is very suitable for use in image reconstruction in the field of X-ray imaging. Particularly when using higher energies or as the consequence of special constraints when X-raying, unwanted image artifacts appear in the reconstruction results. This is a side effect that reduces the quality of the data, but that cannot be avoided with conventional methods. A self-learning algorithm that recognizes the effects of such an unwanted structure and can correct for them while taking into account the prevailing physical and mathematical realities would be a leap forward in quality for nondestructive testing.

Research topic: Known operator learning

One of the core tasks of the new team led by group manager Prof. Andreas Maier will be to confront the challenge of the so-called black box. This refers to the methodologically determined property of a deep-learning-based AI approach, whereby it is not possible for humans to know what decision paths the AI took and how it arrived at its results. Accordingly, it is difficult for people to identify and correct the sources of errors. “If you train AI based on numerous CT scans of the human body, for example, it will begin to interpret certain bodily structures into other entities where they do not exist. This must not be allowed to happen,” explains Professor Andreas Maier, who holds the Chair of Pattern Recognition at Friedrich-Alexander-Universität Erlangen-Nürnberg.

To solve this problem, the researchers are using specific “prior knowledge” in AI so that it can evaluate the plausibility of its work by itself. This can include things like the knowledge of the basic laws of physics, advance knowledge about objects such as CAD data or material data, and traditional signal processing. A known conjunction of information like this is referred to as a known operator. Its use in AI effectively forces the technology to find only such solutions as correspond with the physical and mathematical prior knowledge.

Opening up new areas of application

The new research group’s medium-term goal is to combine these approaches from AI research with other projects at Fraunhofer IIS so as to create synergies for more reliable image interpretation in nondestructive testing and also for applications in the domain of inline testing. “We want to make our research findings available to other project groups across the institute in a spirit of collaboration so that we can jointly open up new areas of application. We also plan to offer licensing models for the technology to our external partners in the future,” Maier summarizes.

 

"We want to make our research findings available to other project groups across the institute in a spirit of collaboration so that we can jointly open up new areas of application. We also plan to offer licensing models for the technology to our external partners in the future."

Prof. Dr. Andreas Maier
Group Manager for Operator-Based Learning