Neural net

[mark mitchell]

Jim wrote:

>  A recent
> problem given in my Artificial Intelligence class was to write a
> three-layer neural network (1 hidden layer) and then write a
> backpropagation routine to train it on handwritten digit data (in
> Scheme, of all programming lanauges!)  Now, In Scheme, I'm probably
> going to get about 50 percent of it done.  In C, I think I could have
> gotten about 70-100 percent of it done.  In Rev, I think I might have
> been able to finish it, as well as making a nice graphical
> representation of the network and a testing routine, almost "for free"

It's funny you mention this as I was JUST about to write and ask 
(following the 'expert system' thread) if someone had implemented such a 
beast yet.  I would like to make a simple neural net with one or two 
layers of hidden units to implement some experiments I'm doing.  It 
doesn't seem like it would be that tough to do in rev/meta.  I was 
thinking of just using lines in a field/variable to represent each node, 
with the rest of the information contained in various hierarchical items 
in the lines, but perhaps there is a better way of doing it or someone 
already has something finished that they could share.  Anyone?

mark mitchell
Japan