Warn on `NaN` loss

Would it not be convenient, particularily for users with less experienced with artificial neural networks, to print a warning when trying to update parameters with a loss of NaN? This could save a lot of wasted resources.