I found the Connectionist Foundations reading to be interesting in that it advocates a completely different approach for emulating human intelligence than the Metacat paper does. The main purpose of the Metacat project is to create a program that can examine and solve analogies by using creativity. While Metacat is practically unique in this fashion, it’s ability to learn is limited to individual trial runs and is unable to use what it’s learned for new, unseen problems. This limitation does not exists in connectionist systems, and in fact the ability to learn throughout several trials is the most significant advantage of neural networks. However, the process of trial and error for training neural networks through back propagation techniques lacks any creativity, and while a trained network can solve new, unseen problems, it lacks the creativity of Metacat. This is because the Metacat program uses a rule set of very abstract relationships that allows it to rationalize its solution in a very human way, and the training process of neural networks only allows these networks to solve a specific type of problem and even then there is no high level rationalization that occurs.
My question is which approach leads to a better understanding of human behavior and thought processes? Is the ability of thinking abstractly and creativity championed by Metacat that leads to a greater level intelligence, or is it instead the ability to learn that exists in connectionist systems?