Solving Inverse Problems With Deep Neural Networks - Robustness Included?

Explore the robustness of deep learning methods in solving inverse problems through this 29-minute talk by Martin Genzel at the Hausdorff Center for Mathematics. Delve into an extensive empirical study examining the resilience of deep-learning-based algorithms against adversarial perturbations in underdetermined inverse problems. Discover findings that challenge previous concerns about instabilities, revealing surprising robustness in standard end-to-end network architectures for tasks such as compressed sensing with Gaussian measurements and image recovery from Fourier and Radon measurements. Gain insights into a real-world scenario involving magnetic resonance imaging using the NYU-fastMRI dataset. Learn about the implications of these results for the reliability of deep learning methods in safety-critical applications, and understand how common training techniques can produce resilient networks without sophisticated defense strategies.