How to create an additional output for ANN where the inputs are considered Not Recognizable?

I know how to implement ANN for pattern recognition in Matlab. I provide input A,B,&C, and train the ANN to recognize A,B,&C.

If I provide input for a totally different class, eg. D the ANN will still try to map it to one of A,B or C, whichever it thinks has the closest similarity to D.

But I don't want it to behave that way, is it possible to create an extra output where the decision is "non of the inputs used during training"?