arxivmath.NAcs.LGmath.AP2026-07-15
Approximation of solutions of parameter-dependent problems by residual neural networks
We develop a convergent scheme to train neural networks involving analytic activation functions based on gradient flows. Convergence properties are guaranteed by Lojasiewicz theory. The main advantage of this approach is its simplicity of implementation. The coefficients of the n…