CRNNdemo

[Jonathan Impett]

Hello,
    In attempting to use the continuous time recurrent network from
neurolib, I notice that the activation of the two nodes in CRNNdemo
arrives at its final value as soon as it enters the loop (by printing
out the values and no. of cycles of the loop). I've tried this with more
complex networks (10+ nodes), and the behaviour is the same.
    If I add <[neura clamp: inp];> (with inp still all at 0.0) into the
stabilisation loop, the values take a little longer to settle, and
arrive at a different attractor, which somehow looks a little more
credible.  In the 2-node demo, this increases the cycles to stabilise
from an invariable 2 to 7-10.
    Can anyone tell me: is it correct that a CRNN should stabilise
immediately (i.e. am I corrupting its behaviour by calling clamp again),
or is this change indeed necessary?
    (subtext - I'm looking for reassurance that I have indeed
<mis>understood)

    My amended chunk of CRNNdemo:

(int loop)

      loop = 0;
      while ((oldOut[0]!= out[0]) && (oldOut[1] != out[1]))
        {
            loop++;
            [ptData resetData];
            for (j = 0 ;j < numNeurons; j ++)
                oldOut[j] = out[j];
            printf("loop %d out[0] %f out[1] %f\n",loop, out[0],
out[1]);
            [ptData addX: out[0] Y: out[1] ];
            [wgtData addX: out[0] Y: out[1] ];
            [neura getActivation: out];
             [neura clamp: inp];    // this is the line I'm tempted to
add (inp is all set to 0.0)
        }

    Thanks,
    Jonathan

--

Jonathan Impett
Faculty of Music
University of Cambridge
West Road
Cambridge CB3 9DP
UK

tel: (+44) (0)1954 250192
address@hidden




                  ==================================
   Swarm-Support is for discussion of the technical details of the day
   to day usage of Swarm.  For list administration needs (esp.
   [un]subscribing), please send a message to <address@hidden>
   with "help" in the body of the message.