Abstract : Inverse problems correspond to the reconstruction of hidden objects from possibly noisy indirect measurements and are ubiquitous in a variety of scientific and engineering applications. Since these problems tend to be ill-posed, and real-world applications are often large-scale, this can be a very challenging task. This minisymposium focuses on recent advances in computationally efficient methods for solving large-scale inverse problems in imaging, e.g., those arising in medical, geophysical and industrial applications, covering topics that include advances in iterative methods, regularization, machine learning and novel applications of the previous.
00255 (1/2) : 3D @E819 [Chair: Malena Sabaté Landman]
[03339] Streaming Methods for Inverse Problems
Format : Online Talk on Zoom
Author(s) :
Eric de Sturler (Virginia Tech)
Abstract : We discuss Golub-Kahan type methods for streaming problems. As big data applications become ever more prominent, in many applications we can only solve problems such as linear or nonlinear regression problems in chunks. Data may come in over a larger span of time and we cannot (or prefer not) to wait until all data is available, the problem may be too large to fit in memory, or data is coming in at a rate that we can use only sampled data and use it in chunks. There is a need for methods that can work efficiently under such conditions. We discuss extensions for GKB-type methods that select and build effective search spaces over multiple subsets of data and/or matrix-blocks to compute accurate solutions with limited memory available. Apart from streaming applications this may also be useful for modern computing architectures that have highly non-uniform memory access.
This is joint work with Julianne Chung, Jiahua Jiang, Misha Kilmer, and Mirjeta Pasha.
[03433] Sequential model correction for nonlinear inverse problems
Format : Talk at Waseda University
Author(s) :
Arttu Arjas (University of Oulu)
Andreas Hauptmann (University of Oulu)
Mikko Sillanpää (University of Oulu)
Abstract : Linear inverse problems are usually solved with first-order gradient methods. For nonlinear problems one must resort to second-order methods that are computationally more expensive. In this work we approximate a nonlinear model with a linear one and correct the resulting approximation error. We develop a sequential method that iteratively solves a linear inverse problem and updates the approximation error. We analyze the sequence theoretically and present numerical results.
[02062] Plants, robots and dynamic tomography
Format : Talk at Waseda University
Author(s) :
Tommi Heikkilä (University of Helsinki)
Abstract : The need for dynamic tomography can arise from many applications, e.g. imaging nutrient perfusion in plant stems for carbon uptake and metabolism studies. While the measurements may be sparse, obtaining the reconstructions can be computationally intensive, since 3D volumes evolving over time leads to 4D tomography. Thus we need fast and efficient methods: our choice are sparse representation systems such as (complex) wavelets and (cylindrical) shearlets, tested on dynamic data from a motorized phantom.
[04602] Deep learning methods for data-driven uncertainty quantification
Format : Talk at Waseda University
Author(s) :
Ling Guo (Shanghai Normal University )
Abstract : In this talk, we will present some recent developments on using Physics-informed neural networks (PINNs) to quantify uncertainty propagation in a unified framework forward, inverse and mixed stochastic problems based on scattered measurements. We will also present generative models for data-driven uncertainty quantification, including physics-informed generative adversarial networks and Normalizing field flows.
00255 (2/2) : 3E @E819 [Chair: Malena Sabaté Landman]
[04412] Exploiting Mixed Precision Arithmetic in Image Reconstruction
Format : Talk at Waseda University
Author(s) :
James Nagy (Emory University)
Abstract : Although some work has been done to exploit mixed precision computations for inverse problems arising in image processing, most previous work focuses on extended precision to avoid the influence of rounding errors. We consider a different perspective: because we cannot expect to precisely know data, we develop and analyze solvers that can take advantage of low precision speed of modern computer architectures, and which can be used in a variety of imaging applications.
[03595] Image Quality Assessment for Reconstruction Algorithms
Format : Talk at Waseda University
Author(s) :
Anna Breger (University of Cambridge)
Abstract : Assessing digital image quality (IQ) is of high importance in numerous applied research fields, such as image acquisition, reconstruction or processing. Automated evaluation is needed since manual evaluation is too time-consuming and expensive for huge data sets and, furthermore, may be biased or introduce inconsistencies. In this talk I will give a short introduction to IQ assessment and provide examples of failure when applying standard full reference IQ measures in medical imaging reconstruction tasks.