This course focuses on exploring the robustness of deep-learning-based algorithms for solving underdetermined inverse problems, specifically in image reconstruction tasks. The main goal is to conduct an empirical study on the reliability of deep neural networks when faced with adversarial perturbations in the input domain. The course covers compressed sensing with Gaussian measurements, image recovery from Fourier and Radon measurements, and a real-world scenario for magnetic resonance imaging. Participants will learn about computing adversarial perturbations that maximize reconstruction error and discover the resilience of standard end-to-end network architectures against statistical noise and adversarial attacks. The intended audience for this course includes individuals interested in deep learning, image processing, and inverse problems.
Solving Inverse Problems With Deep Neural Networks - Robustness Included?
Hausdorff Center for Mathematics via YouTube
Overview
Syllabus
Martin Genzel: Solving Inverse Problems With Deep Neural Networks - Robustness Included?
Taught by
Hausdorff Center for Mathematics