While it may not be necessary to understand why a certain decision was made for an AI system to perform adequately, it is necessary to build trust in those systems. Thus, the transparency and explainability of AI systems are one of the biggest research topics right now. Most AI approaches start with data from which information for action is derived. However, the conditions under which such data is collected is rarely considered, although a closer look at the sensors could lead to more robust and widely applicable algorithms. Sensors and their working principles are well understood because scientists have been building detailed physical models for decades, and calibration of such sensors is well established in science, industry. Therefore, physical and data-based models could be merged to design revolutionary AI systems. This field of AI system design is called “Sensor AI”. 

Robonaut 2, a dexterous, humanoid astronaut helper. Image Credit: NASA/JSC

The paper proposes two directions of further research for the sensor artificial intelligence: 

High-level data processing onboard satellites: due to the discrepancy between the volume of acquired data and downlink capabilities there is a need for a drastic data reduction that cannot be provided by ordinary compression approaches. Therefore, high-level data processing onboard satellites will be required.  Overall optimization of current artificial intelligence systems: Spaceborne instruments can be built and operated in an optimal way only if they are considered as a part of a very complex system including the object of observation, the environment, the sensor itself and the data processing. Such complex systems cannot be modelled in purely a physical way, and so AI methods can help to approximate these models.

 The sensor AI promises not only to improve the transparency of current AI practices but also expand the range of applications of those AI systems.

Link to slides

Main Reference   Sensor Artificial Intelligence and Its Application to Space Systems

Additional References   NASA's Fields of Application for Artificial Intelligence

  • No labels