the neural network must be recursive, not top to bottom like current ANNs. The network outputs stops when ready on its own, not when it gets to the last layer. You would need a new definition of done. One definition is that the inputs match the output in some way.