These figures were compared to the same measurements for a standard backpropagation
network (with the same architecture as the final-storage area of the pseudo-recurrent
network) that had first learned the background set of patterns and then learned the new
quasi-contradictory set of mushrooms.