Shen and the ANN

[Antti Ylikoski]

I have been working on the artificial neural net backpropagation

algorithm for the Shen system.  The backpropagation ANN is (one of)

the workhorse(s) of the neural net business (though newer ANN schemes

exist.)

I have been collecting various improvements to the basic

backpropagation ANN algorithm.  Then, I came across the below paper:

Alpsan-Towsey-Ozdamar-Tsoi-Ghista: Are Modified Back-Propagation

Algorithms Worth the Effort?  ICNN '94.  (not exactly very modern)

In short, the authors answer, mostly not.

I compared, in an email, the business of making a truly good ANN

package with Shen, to building a truly good car in the year 1925.  Not

impossible, definitely not, but a pretty nontrivial job.

After a number of changes to the design, and the above observation,

the backpropagation ANN is being debugged, and coming Right Soon Now;

in the form of the basic backpropagation, well, with the momentum

constant.

I have written this mostly to be able to avoid repeated work: I know

that other individuals in this group are doing ANN work.

A. J. Y.