I was surprised to see that the regularization term is divided by n_samples. This is not standard.

https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/neural_network/multilayer_perceptron.py#L125

This doesn't seem to correspond to the objective documented in
http://scikit-learn.org/dev/modules/neural_networks_supervised.html