Can fit neural networks with up to two hidden layer and two different error functions. Also able to handle a weight decay. But just able to compute one output neuron and very slow.