Alternating mixed-integer programming and neural network training for approximating stochastic two-stage problems