Current browse context:
math.NA
Change to browse by:
References & Citations
Mathematics > Numerical Analysis
Title: Deep neural networks can stably solve high-dimensional, noisy, non-linear inverse problems
(Submitted on 2 Jun 2022 (v1), revised 13 Jun 2022 (this version, v3), latest version 20 Oct 2023 (v5))
Abstract: We study the problem of reconstructing solutions of inverse problems with neural networks when only noisy data is available. We assume that the problem can be modeled with an infinite-dimensional forward operator that is not continuously invertible. Then, we restrict this forward operator to finite-dimensional spaces so that the inverse is Lipschitz continuous. For the inverse operator, we demonstrate that there exists a neural network which is a robust-to-noise approximation of the operator. In addition, we show that these neural networks can be learned from appropriately perturbed training data. We demonstrate the admissibility of this approach to a wide range of inverse problems of practical interest. Numerical examples are given that support the theoretical findings.
Submission history
From: Philipp Petersen [view email][v1] Thu, 2 Jun 2022 08:51:46 GMT (205kb,D)
[v2] Tue, 7 Jun 2022 15:45:13 GMT (211kb,D)
[v3] Mon, 13 Jun 2022 09:50:25 GMT (211kb,D)
[v4] Tue, 21 Jun 2022 07:57:09 GMT (212kb,D)
[v5] Fri, 20 Oct 2023 14:06:30 GMT (193kb,D)
Link back to: arXiv, form interface, contact.