[pdp-discuss] TesselPrjnSpec usage

[Richard Jones]

I've switched over to rbp++ and got it figured out.  Thanks.

BTW, when looking for a NN tool to play with earlier this year I read
a review of PDP++ somewhere that said basically "good features, bad
support".  I've found that the mailing list support has been prompt,
the documentation is thorough, and the code is organized and
well-written.



On 10/14/06, Randall C. O'Reilly <Randy.OReilly at colorado.edu> wrote:
> unless you are using rbp, you can't have a self-projection in bp -- it is
> feedforward only..
>
> You did set the init_wts flag in the spec, and also set weight values in the
> individual tessel elements, right?  That should work, even if it doesnt' make
> sense to do it in bp..
>
> Those functions are in the "Actions" menu of the tessel edit dialog..
>
> - Randy
>
> On Friday 13 October 2006 16:22, Richard Jones wrote:
> > Hi,
> >
> > I'm trying to create a self-projection for a hidden layer using a
> > TesselPrjnSpec.  I'd like to have a RF pattern along the lines of the
> > one in the book's "v1rf" project in chapter 8.  However, I'd like each
> > of my units to try and inhibit his near neighbors.
> >
> > bp++ allows me to assign the TesselPrjnSpec to the self-projection in
> > question, and I can edit its properties.  However, after applying and
> > doing a "Connect All" (or any of the other init buttons here and
> > there) no non-zero weights are created.  It doesn't seem to matter
> > which params I set in the properties editor.   If I assign other
> > PrjnSpec types (e.g., PolarRndPrjnSpec) I get non-zero weights.
> >
> > I notice that in leabra++ with the v1rf project the TesselPrjnSpec
> > does work.  Does it perhaps require that I use a LeabraConSpec instead
> > of a BpConSpec?
> >
> > Lastly, the online help makes mention of functions associated with the
> > TesselPrjnSpec (e.g., MakeEllipse(int half_width, int half_height, int
> > ctr_x, int ctr_y)), but it's not clear whether this is something
> > invoked automatically from the GUI, or if it's something that needs to
> > be invoked from some code somewhere.
> >
> > How do I get that nice guassian cluster of weights around my sending unit?
>


-- 
Richard Jones
dickjr at gmail.com