One of the objections of much of the work that has been done in the area of catastrophic
interference is that the patterns to be learned are tailored to produce the desired phenomenon.
For this experiment, we will consider data from the UCI mushroom database (Murphy & Aha,
1992). The reason that it is important to use real databases is that these presumably code for
“real world” regularities. It can be shown (Hetherington & McRae, 1993) that when a neural
network is pretrained on a random sample of data from the domain in which it will be
operating, catastrophic interference can be eliminated