This is a paper I wrote many years ago that appears to be a reasonable way to instantiate a neural network with prior knowledge. The method is a modification of the more ubiquitous KBANN algorith and appears to outperform it in terms of training time.

The power of this algorithm comes from decoupling the structure of the newtork to the domain knowledge. It should exhibit reasonably good behaviour even if the prior knowledge is reasonably uncertain. The training time is also much reduced since the network only contains a single layer of hidden units - thus its name: Shallow-KNANN.

Download the postscript version here: SKBANN Postscript
Download the pdf version here: SKBANN PDF

The PostScript version looks best. Linux people can use ghostview or something. For Windows a very good Postscript interpreter and viewer is available at the AFPL site.