Question about presenting target output

[alkim@cattell.psych.upenn.edu (Albert E Kim)]

Dear PDP++ list: 

I'm having a problem creating a network in which the target pattern is 
presented to a layer of units that is not the last layer in the
network. 

I'm trying to create a network with simple recurrent feedback from the
output layer.  That is, at each time step, the network should copy the
current output pattern to a "context" layer.  At the next time step,
the copied output pattern should be passed to the hidden layer.  This
is just like the SRN example in pdp++, except the copying is
Output-to-Copy instead of Hidden-to-Copy.  Architecture looks like
this (NOT SHOWN are OUTPUT-to-COPY connections):  

	COPY_OF_PREVIOUS_OUTPUT	 ==>	HIDDEN	==OUTPUT	
	INPUT			 ==>

I created a new Copy layer, which is numerically AFTER the Output
Layer, so that the Output Layer should update its activations before
they are copied by the Copy layer.  However, this makes the Copy layer
the LAST layer in the network, which makes it the layer that gets
external targets.  Can anybody tell me how to make pdp++ present the
target patterns to the Output layer instead of the Copy layer and
backpropagate error from there?? 


Thanks in advance,
Al Kim
Research Fellow,
Moss Rehabilitation Research Institute