A Deep Learning Network for the Natural Sciences

Project funded by PIER         

In many scientific disciplines, the extraction of three dimensional spatial features from either sparse or incomplete data is a key technology for the interpretation of experimental data. In some cases even time information as fourth dimension is essential. This is especially true for the PIER research areas, where astrophysical images of galaxies, photon scattering data at the molecular level and particle physics images are investigated in Hamburg. A common denominator for all these cases is missing information from at least one dimension and the challenge to invert the problem from image space to the space of interest. This together with the need to process very large sets of images leads to major intellectual and computational problems. 



The aim of this project is to initiate a core group of scientists from different fields which has to potential and the clear vision to develop into a network for deep learning algorithms based on machine learning and even more generally for artificial intelligence in Hamburg.

For the core group, we will work on 

  •  Dedicated hardware platform enabling state of the art machine learning

  • Image processing and data preprocessing

  • Key technologies in network optimization

  • Understanding of network capabilities using deconvolution techniques

  • Generative Adversarial Networks and prior knowledge in science

  • Expert knowledge in deep learning algorithms

  • Regression techniques for model parameters in natural science 

Project Partners:

Prof. Henry Chapman, CFEL
Prof. Simone Frintrop, Fachbereich Informatik, Universität Hamburg
JunProf. Gregor Kasieczka, Institut für Experimentalphysik, Universität Hamburg

Prof. Peter Schleper, Institut für Experimentalphysik, Universität Hamburg 

Below are a list of topics that we are working on in my group:
Radio Galaxy Classification using Deep Learning (LOFAR data; Julia ML libraries, image augmentation)