On the convergence of a data-driven regularized stochastic gradient descent for nonlinear ill-posed problems (Q7013083)

In this interesting paper, the authors are concerned with the numerical study of systems of operator equations of the form \(F_i(x)=y_i\), \(i=1,\dots,n\) where each \(F_i:D(F_i)\to Y\) is a nonlinear mapping with domain \(D(F_i)\subset X\) and \(X, Y\) are Hilbert spaces. Such nonlinear inverse problems arise for example in inverse scattering and electrical impedance tomography. This paper studies such representations according to certain prior knowledge of the problem given by certain algorithms in the paper which incoporate prior knowledge for the problem. Stochastic gradient methods as first proposed by \textit{H. Robbins} and \textit{S. Monro} [Ann. Math. Stat. 22, 400--407 (1951; Zbl 0054.05901)] play an important role in the paper. Indeed, stochastic gradient descent is a promising method for solving large-scale inverse problems, due to its excellent scalability with respect to data size. The authors, more precisely, study a data-driven regularized stochastic gradient descent for the efficient numerical solution of a class of nonlinear ill-posed inverse problems in infinite dimensional Hilbert spaces. At each step of the iteration, the author's interesting construction selects one equation from the nonlinear system combined with a corresponding equation from the learned system based on training data to obtain a stochastic estimate of the gradient and then performs a descent step with the estimated gradient.\N\NThe authors establish rigorously a regularizing result and provide numerical examples.