Approximating Langevin Monte Carlo with ResNet-like Neural Network architectures