We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
A Loss Landscape Perspective and Simulations for Imaging Inverse Problems based on AI and Neuron Network Training Method.
- Authors
Mingyong Zhou
- Abstract
The purpose of Imaging Inverse problems are to recover original true signals/images from observations by plethora physical methods. The recovered signals/images are measured by appropriate quantity and quality measures. Given a specific physical process, imaging is classically a mathematical algorithm design to estimate the original signals/images. In this paper starting from inverse imaging problem of Electrical Impedance Tomography (EIT), we will address several aspects of loss landscape and its regularization properties of a neuron network solution for imaging inverse problems. Loss landscape and regularization strategy for imaging inverse problems is first introduced in I), and is followed by II) Sampling from unbounded landscape to bounded discrete. In III), an asymptotic approximation property when neuron network width goes to infinity is analyzed. Complexity of training and approximation error is discussed in IV). In V), case study for indicator function approximation and application in imaging EIT by neuron network is highlighted. Finally in VI), discussions will be devoted to the limitations of regularization capability of neuron network as a new tool, especially in medical imaging applications.
- Subjects
ARTIFICIAL intelligence; INVERSE problems; ELECTRICAL impedance tomography; NEURONS; APPROXIMATION error
- Publication
International Journal of Simulation -- Systems, Science & Technology, 2024, Vol 25, Issue 1, p6.1
- ISSN
1473-8031
- Publication type
Article
- DOI
10.5013/IJSSST.a.25.01.06