Re: Problems, setting parameters for single units

["Randall C. O'Reilly" <oreilly@grey.colorado.edu>]

Mathias Weigt <weigt@uni.de> writes:
> But now I wanted to make some differentiations on how the input-parameters 
> are treated (e.g. that I can exclude one or many parameter, without 
> redesigning the input-layer and the event-set).

by parameter, I'm inferring that you mean activation of an input unit?

> 1. Manipulate on the unit object: Creating new child of the Unit-spec and 
> setting the activation range to zero (the unit should not get any activation 
> any more). Then applying this to a single unit on the input layer.
> Result: It is ignored.

this is a good idea, but yes, the activation range is ignored for
hard-clamped inputs.. one idea is that you could change the units to
use soft clamping -- but this is not avail in standard bp++ which I'm
assuming you're using (it is only in rbp or cs where units update over
time).

> 2. Creating a new Con-Spec with Uniform weights (mean=0, var=0). This should 
> have the advantage, that I can differentiate for which receiver the weight 
> should be zero. 
> Results: when applied to a single unit, it is ignored
> when applied on the Projection the whole layer is treated...
> Then I read further in the documentation and found, that ConSpec it always 
> used for the whole unit-group and my input-layer was of one group (z=1).

what this means is that there is one conspec for an entire group of
*connections* (not units) associated with an individual unit (i.e., a
con_group on a unit).  So, you could only change all of the input
weights for a given unit to zero, which wouldn't help too much.

The bottom line is that this manipulation is designed in pdp++ to be
done in the environment, not in the network.  You could for example
use the pattern flags in the environment to pick certain inputs to not
be presented -- if you use the global flags associated with the
pattern spec, you only need to set these once and it will affect all
of the events, without having to change the event patterns
themselves.  This is what I would recommend.

				- Randy