Can Imaging Still See the Unseen in the Age of AI? Implications for Scientific Discovery

Abstract

For centuries, imaging technology has been a major enabler of scientific discovery, providing novel windows into unseen worlds. Classically, such systems were designed to be as objective as possible, based on well-understood operating principles with well-characterized limitations. However, this has changed in recent years with the increasing availability of big data and revolutionary progress in machine learning and artificial intelligence. Indeed, learning-based computational imaging systems have now become regarded as the current state-of-the-art. That said, it is important to recognize that learning-based approaches are often biased towards the training distribution, and that our definition of the state-of-the-art rests on a reorientation of our values compared to the past. The presence of bias has potentially profound implications for scientific imaging, since it may not be clear whether a reconstructed image is truly a reliable representation of physical reality. In this work, we describe a new approach for testing imaging system performance in novel regimes. Our testing reveals that some popular computational imaging approaches have limited performance in novel regimes, warranting caution about how and when they should be used.

Ähnliche Arbeiten