Constrained Neural Networks

• Applicant: Professor Dr. Michael Möller
• Subject Area: Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
• Term: from 2020 to 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 448537382

• Final Report Year: 2025

Final Report Abstract

The so-called artificial intelligence (AI) is technically realized primarily through artificial neural networks. These are complex parameterized functions for which one aims at adjusting the parameters so that the functions produce a desired output for a given input. In the simplest case, this adjustment is achieved by using (training) examples, i.e., pairs of inputs and desired outputs specified by humans. Despite the incredible progress made over the past 10 to 15 years in training neural networks on large datasets, the trained networks often exhibit surprising instability. Even if all the given training data fulfill a specific property or are generated according to a certain physical relationship, there is no guarantee that the predictions made by neural networks will respect such relationships. This can lead to significant problems, particularly in safety-critical applications: even well-known conditions that can be precisely described using mathematical (in)equalities cannot be provably guaranteed. For example, in the reconstruction of medical images in computed tomography using neural networks, it is not inherently ensured that the image produced by the AI actually corresponds to the acquired data. For this reason, this project focuses on technical approaches to provably ensure that known constraints are respected in machine learning methods. The project’s focus is on applications in the field of machine vision, such as the automatic segmentation of images into specific regions, the automatic solution of correspondence problems on three-dimensional shapes, and the reconstruction of images from measurement data that contain implicit information about the desired image.

Publications

• Learning or modelling? an analysis of single image segmentation based on scribble information. In International Conference on Image Processing (ICIP), 2021
  Hannah Dröge & Michael Moeller
  
• Physical Representation Learning and Parameter Identification from Video Using Differentiable Physics. International Journal of Computer Vision, 130(1), 3-16.
  Kandukuri, Rama Krishna; Achterhold, Jan; Moeller, Michael & Stueckler, Joerg
  
• Explorable data consistent ct reconstruction. In BMVC, 2022
  Hannah Dröge, Yuval Bahat, Felix Heide & Michael Moeller
  
• Non-smooth energy dissipating networks. In International Conference on Image Processing (ICIP), 2022
  Hannah Dröge, Thomas Möllenhoff & Michael Moeller
  
• Evaluating adversarial robustness of low dose ct recovery. In Medical Imaging with Deep Learning (MIDL), 2023
  Kanchana Vaishnavi Gandikota, Paramanand Chandramouli, Hannah Dröge & Michael Moeller
  
• Implicit representations for image segmentation. In Conference on Neural Information Processing Systems Workshops, 2023
  Jan Philipp Schneider, Mishal Fatima, Jovita Lukasik, Andreas Kolb, Margret Keuper & Michael Moeller.
  
• Kissing to find a match: Efficient low-rank permutation representation. In Conference on Neural Information Processing Systems (NeurIPS), 2023
  Hannah Dröge, Zorah Lähner, Yuval Bahat, Onofre Martorell Nadal, Felix Heide & Michael Moeller.
  
• On the Confluence of Machine Learning and Model-Based Energy Minimization Methods for Computer Vision. PhD thesis, University of Siegen, 2023
  Hannah Droege