Introduction to Connectionist Modelling of Cognitive Processes
Peter McLeod, 1998

The Two Parts of this Book
The first part of this book is a sort of introduction to neural networks. I say that with reservation because it is basically a non-mathematical treatment of what I consider to be an essentially mathematical topic. For example, not even the basic process of backpropagation training is presented in its full detail. The formulas given are incomplete and inadequate to understand or implement a backprop-trained network. On the other hand, the treatment is fairly broad, covering Hebb networks, as well as recursion.

The second part of the book has much more to say. In showing how neural nets might be applied to implement some of the functions of the brain, several such brain components are described in enough detail to be very educational. Of particular interest, I found the hippocampus discussion in chapter 13, clearly the work of Edmund Rolls, to add new insights into the organization and function of this vital brain system.

The third feature of the book is a program which implements many of the neural systems discussed throughout the book. The tlearn program, available on the diskette provided, includes files to set up and train the network and to display the results in an understandable manner. Although billed as a Windows95 program, it would not run on my Win95 system. A pop-up window says that one of the WindowsNT DLLs is not available.

Attractor Networks
I have a bit of a terminology problem with attractors. I find the
treatment here to be correct, but not adequately clear. The problem is this:

The action of a recurrent network in converging upon the final output involves a sequence through a series of states. It is the operation of the network state stepping sequentially toward the final result that is described by the term attractor. This developing sequence is most clearly seen when shown as a phase diagram. The diagrams here almost make the right point, but, in my opinion, do not adequately present the idea of the phase space. This issue would all be inconsequential were it not for the confusion which arises in thinking of any pattern recognizer as an attractor. I find an unfortunate tendency in some other places for the term to be devalued by its use in this latter sense.

Top of Page | ICMCP Summary | Sort by Topic | Sort by Title | Sort by Author