Author |
Message |
JLS
Joined: Nov 05, 2005 Posts: 490 Location: Czech
Audio files: 30
G2 patch files: 316
|
|
Back to top
|
|
|
blue hell
Site Admin
Joined: Apr 03, 2004 Posts: 24079 Location: The Netherlands, Enschede
Audio files: 278
G2 patch files: 320
|
Posted: Wed Sep 30, 2015 4:23 am Post subject:
|
|
|
Thanks :-)
Will look into it. _________________ Jan
also .. could someone please turn down the thermostat a bit.
|
|
Back to top
|
|
|
JLS
Joined: Nov 05, 2005 Posts: 490 Location: Czech
Audio files: 30
G2 patch files: 316
|
|
Back to top
|
|
|
blue hell
Site Admin
Joined: Apr 03, 2004 Posts: 24079 Location: The Netherlands, Enschede
Audio files: 278
G2 patch files: 320
|
Posted: Sun Oct 11, 2015 12:43 pm Post subject:
|
|
|
I could see some use for neural networks, and in fact had started some work on them before ... with a module looking like:
Internally this uses the FANN ( http://leenissen.dk/fann/wp/ ) code for network calculation. This module can process feedforward multi-layer networks with 12 input and 12 output neurons. The number of neurons in the internal layers is pretty much arbitrary.
When i understand it right a perceptron is a feedforward network, and as such the FANN library could implement it.
The reason that I never published this module is .. that I see no handy way for training networks. Training would have to be an offline, non-realtime operation, and so it would need a separate program, or a separate (learning) mode of Wren to perform the calculations.
For the network to do useful audio related things you would probably want to train it on audio data, so to have learning as some separate mode in Wren seems to make most sense then.
Anyway, sofar I have not been able to come up with a solution for this issue.
Re. the Hopfiled network, as far as I can tell the FANN library is not able to handle such networks, so I'd have to implement that from scratch, or possibly using some other toolkit. Still, there would be the same issue there ... what would be a good way to train the network.
Training requires two things to be present, being 1) a set of training data and 2) some desired 'outcome' for that data.
I have been experimenting a bit with this for a different project where I had a couple sensors from which I wanted to draw some conclusions. For that purpose I then wrote a graphing program in which I could just draw the desired outcome and then could make the network learn from 1) the measured data and 2) the drawn graph. This did give some reasonable output for the problem then (although sales thought 90% accurate results to be not good enough). Actually .. I did use a modified (non-realtime) version of Wren for this project to do the calculations for me (and that is where this neuron module actually came from)
Anyway, I think a network module would need extra tools, like the one mentioned above, before they would be useful.
Then again, this all is my limited vision _________________ Jan
also .. could someone please turn down the thermostat a bit.
|
|
Back to top
|
|
|
JLS
Joined: Nov 05, 2005 Posts: 490 Location: Czech
Audio files: 30
G2 patch files: 316
|
|
Back to top
|
|
|
JLS
Joined: Nov 05, 2005 Posts: 490 Location: Czech
Audio files: 30
G2 patch files: 316
|
|
Back to top
|
|
|
|