I tried Kohonen for classifying timbres once and though I'm certainly not an expert on it Ben I seem to remember that it works within a threshold value. Input vector is just a _point_ in an n-dim space with a "sphere of influence" around it. If an output vector is within this threshold area then the more input stimulus you present that match a vector falling inside the threshold the more that output is weighted. Eventually that output vector will fire with probability= 1 (absolutely certain). Further input vectors, no matter how many you give, won't affect output vectors outside the threshold. It's a long time ago, so excuse me if I'm talking rubbish.
On Fri, 12 Oct 2007 11:27:15 -0700 "B. Bogart" bbogart@goto10.org wrote:
Hey all, Johannes in particular,
According to Medler: http://neuron-ai.tuke.sk/NCS/VOL1/P3_html/node28.html#SECTION000450000000000...
"...Consequently, more frequently occurring stimuli will be represented by larger areas in the map than infrequently occurring stimuli."
Which leads me to believe if you present a SOM with the same pattern, the number of nodes that fire should increase proportional to the frequency of the pattern.
Problem is I can make ann_som do that, not matter what rules I specify (using the regular decreasing learning rate).
If I send the same pattern to the som 10000 times, only one node will fire for that duration.
Am I misinterpreting the text?
Also by default does ann_som using time-seeded random numbers or regular random numbers? I ask because it appears the same node always wins at the start... (the biggest number).
Is there a FANN mailinglist for the discussion of such issues?
(http://leenissen.dk/fann/ is timing out for me.)
Thanks, B.
PD-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list