[pdp-discuss] relation layer

[Randall C. O'Reilly]

Matt,

You may need to play with the lrate, number of hidden units and other params 
-- backprop does not always learn.  I did a lot of work with the family tree 
model, which has a somewhat similar structure to your task, and it took a lot 
of work to get backprop to work -- you can look in my phd thesis on my 
website for details on that: 
http://psych.colorado.edu/~oreilly/pubs-online.html  Also, make sure your 
layers are in the right feedforward order!

- Randy

On Friday 23 December 2005 13:05, Matty Dee wrote:
> Hi all. I am new to this list and to PDP++ in general. This is the
>  first network I've ever worked on so I don't know very much at this
>  point. The following link points to a screenshot from PDP++ (using
>  BP++) and should by itself explain what I am trying to achieve:
>
>  http://i32.photobucket.com/albums/d43/deebaysir/net.jpg
>
>  The network doesn't seem to be able to learn the difference between the
>  relations "has" and "can". As far as I've read on this matter, the
>  network should quickly learn to turn off attributes that are not
>  related to the proper relation unit... it doesn't.
>
>  It seems to want to balance out the error by turning on all target units
> related to the character regardless of the relation ("has" or "can").
>
>  The following link points to a graph of the learning curve:
>
>  http://i32.photobucket.com/albums/d43/deebaysir/graph.jpg
>
>
>  I'm hoping there is something simple that I am missing here, since it
>  seems that this should be a basic network. Any help would be very
>  much appreciated.
>
>  Thank you.
>
>  Matt