This is the example used in the tutorial. It is very easy to memorize the training set, but can the network get the testing set correct even though some bits have been changed?
Try making a new network that doesn't have any hidden layers. You should be able to do it in a single command using addNet.
Now try using some weight decay, like maybe around 0.001. Does that help? Perhaps not. Oh well, it's a silly task anyway.