DeepSAM: Machine Learning under Symmetries with Applications in Astronomy and Medicine

While at vastly different scales and proportions, many scientific fields face similar problems. One such problem, shared by both astronomy and medicine, is the proper automated treatment of image-like structures. Due to the vast amount of available data, deep learning methods are predestined to tackle this task. Although much advance has already gone into this field, most attempts still fail to properly leverage the symmetries of the underlying systems. For example, the orientation of both cells and galaxies has no inherent meaning and should therefore not be taken into account when building proper machine learning models.

In this project, we aim to develop novel deep learning methods, that are well equipped to treat these problems in a way that respects the intricacies of the data, while allowing a level of interpretability that is crucial for usage in the natural sciences. We intend to deploy such methods for the use in both astronomy and medicine.

Contributors:

Kai Polsterer(AIN)

Vincent Heuveline (DMQ)

Members:

Romain Chazotte