Approximation of solutions of parameter-dependent problems by residual neural networks

Ana Carpio

We develop a convergent scheme to train neural networks involving analytic activation functions based on gradient flows. Convergence properties are guaranteed by Lojasiewicz theory…