From c17e6115bda2e692dab2c47dc0be6e343741bc61 Mon Sep 17 00:00:00 2001 From: Guenter Geiger Date: Tue, 19 Nov 2002 12:02:33 +0000 Subject: added doc directory svn path=/trunk/; revision=220 --- externals/build/doc/ann/som.pd | 114 +++++++++++++++++++++++++++++++++++++++++ 1 file changed, 114 insertions(+) create mode 100644 externals/build/doc/ann/som.pd (limited to 'externals/build/doc/ann') diff --git a/externals/build/doc/ann/som.pd b/externals/build/doc/ann/som.pd new file mode 100644 index 00000000..9f039792 --- /dev/null +++ b/externals/build/doc/ann/som.pd @@ -0,0 +1,114 @@ +#N canvas 50 -127 640 687 10; +#X msg 131 495 print; +#X msg 132 528 new 5 8 8; +#X msg 127 99 init; +#X msg 128 274 train; +#X msg 129 296 test; +#X msg 128 387 write; +#X obj 70 559 ann_som 4 9 10; +#X msg 70 49 1 0 0 1; +#X msg 70 68 0 1 0 1; +#X msg 70 87 2 1 0 0; +#X msg 128 118 init 0.5; +#X msg 128 138 init 1 0.5 0 0.5; +#X text 234 101 init all weights with "0"; +#X text 235 120 init all weights with "0.5"; +#X text 235 137 init weights for each sensor; +#X msg 128 163 learn 1; +#X msg 128 197 learn 1 0.9 0.1; +#X text 226 163 set learning rate to 1; +#X msg 128 180 learn 0.5 0.999; +#X text 227 179 set learning rate to 0.5 and factor to 0.999; +#X text 227 197 set learning rate to 1 \, factor to 0.9 and offset to 0.1; +#X msg 128 214 neighbour 1; +#X msg 128 231 neighbour 0.5 0.999; +#X msg 128 248 neighbour 1 0.9 0.1; +#X text 248 215 set neighbourhood to 1; +#X text 249 231 set neighbourhoodto 0.5 and factor to 0.999; +#X text 249 249 set neighbourhood to 1 \, factor to 0.9 and offset to 0.1; +#X text 180 269 set som to "train" mode (learn from sensor-input and output winning neuron); +#X text 179 291 set som to "test" mode (output winning neuron for sensor-input \, but do not learn !); +#X msg 129 328 rule INSTAR; +#X msg 129 345 rule OUTSTAR; +#X msg 129 362 rule KOHONEN; +#X text 218 327 learn with IN-STAR rule; +#X text 219 345 learn with OUT-STAR rule; +#X text 219 362 learn with KOHONENrule; +#X msg 128 405 write mysom.som; +#X msg 129 429 read; +#X msg 129 447 read mysom.som; +#X text 156 68 present various data to the SOM; +#X text 203 495 for debugging; +#X text 207 530 create a new SOM with 8x8 neurons \, each having 5 sensors; +#X text 204 561 create a new SOM with 9x10 neurons \, each having 4 sensors; +#X floatatom 70 614 4 0 0; +#X text 113 618 winning neuron; +#N canvas 13 0 889 630 SOMs 0; +#X text 76 27 SOM :: Self-Organized Maps; +#X text 55 53 SOMs are "Artificial Neural Networks" \, that are trying to learn something about the data presented to them without a supervisor/teacher.; +#X text 59 118 in short:; +#X text 120 119 the neuron \, whose weight-configuration matches the presented data best is the winner (its number (counting from the lower-left corner) is sent to the output); +#X text 121 163 to match the data better the next time it is presented \, the weights of the winning neuron are adjusted.; +#X text 121 188 the weights of the neurons neighbouring the winner are adjusted to match the data too \, but not so strong as the winner's.; +#X text 121 276 lr(n+1)=lr(n)*factor; +#X text 275 277 learning_rate=lr+offset; +#X text 121 289 nb(n+1)=nb(n)*factor; +#X text 275 290 neighbourhood=nb+offset; +#X text 121 230 both neighbourhood and learning-rate (==amount of how much the weights of the winner (and \, proportional \, the weights of the neighbours) are adjusted) are decreasing recursively with time.; +#X text 119 319 thus you will sooner or (most of the time) later get a "brain map" \, where similar inputs will activate neurons in specifique regions (like there are regions for seeing and regions for hearing in our brains); +#X text 97 381 there are various rules \, how to re-adjust the weights of the neurons : in-star \, out-star and kohonen (maybe there are others \, but these i found in literature); +#X obj 607 220 +; +#X text 640 182 ...; +#X obj 579 185 * \$1; +#X obj 607 185 * \$2; +#X obj 670 185 * \$0; +#X obj 579 128 unpack 0 0 0 0 0; +#X text 602 111 n sensors; +#X text 705 186 weights 1 to n; +#X obj 579 90 inlet; +#X obj 607 288 outlet; +#X text 594 62 a neuron; +#X text 566 307 the neuron with the highest weighted sum; +#X text 567 318 matches best and is therefore the winner; +#X text 53 452 notes:; +#X text 101 453 each neuron of the SOM has n sensors. you have to present a list of n floats to the SOM to make it work; +#X text 102 482 you should init the weights for each sensor with the expected mean of the sensor values before you start training to get best and fastest results; +#X text 55 87 they were first proposed by the Finnish scientist T.Kohonen in the 80ies (i think).; +#X text 98 543 if you have no clue \, what's this all about \, maybe you do not need SOMs (which i doubt) or you should have a look at; +#X text 118 577 http://www.eas.asu.edu/~eee511; +#X text 118 591 http://www.cis.hut.fi/projects/ica; +#X connect 13 0 22 0; +#X connect 15 0 13 0; +#X connect 16 0 13 0; +#X connect 17 0 13 0; +#X connect 18 0 15 0; +#X connect 18 1 16 0; +#X connect 18 4 17 0; +#X connect 21 0 18 0; +#X restore 535 44 pd SOMs; +#X text 81 13 ann_som :: train and test Self-Organized Maps; +#X obj 73 660 ann_som test.som; +#X text 211 664 load a SOM-file; +#X connect 0 0 6 0; +#X connect 1 0 6 0; +#X connect 2 0 6 0; +#X connect 3 0 6 0; +#X connect 4 0 6 0; +#X connect 5 0 6 0; +#X connect 6 0 42 0; +#X connect 7 0 6 0; +#X connect 8 0 6 0; +#X connect 9 0 6 0; +#X connect 11 0 6 0; +#X connect 15 0 6 0; +#X connect 16 0 6 0; +#X connect 18 0 6 0; +#X connect 21 0 6 0; +#X connect 22 0 6 0; +#X connect 23 0 6 0; +#X connect 29 0 6 0; +#X connect 30 0 6 0; +#X connect 31 0 6 0; +#X connect 35 0 6 0; +#X connect 36 0 6 0; +#X connect 37 0 6 0; -- cgit v1.2.1