Neural Networks in Max

Apr 29, 2013 at 12:54am

Neural Networks in Max

Hello,
Does anyone know of a way to implement a neural network in Max? I have created and trained a network in Matlab and I would like to use it in a Max patch. What I need is an object that takes the number of inputs, outputs, hidden layers, neurons, and a big list of weights as parameters. Then I could pass it a list as input and it would return the output. Does this exist?
-Adam

#67991
Apr 29, 2013 at 5:49am

A couple of post to get you started…

http://cycling74.com/forums/topic.php?id=16099

http://cycling74.com/forums/topic.php?id=27805

nicolas danet will probably chime in at some point his object Tralala is here

https://github.com/nicolasdanet/Max/tree/master/Tralala/External

The help file 21.Learn for Tralala might be what your looking for.

#244325
Apr 29, 2013 at 6:43am

Hi,

1. You will not find anything related to NN in my repository. The objects that i have done about NN are not avalaible anymore. I started to learn programming while i developped them and the source code was awfull. I decided to throw them. But anyway at this time it was not possible to do what you expect (set the weights from a list).

2. For my object [Tralala] i kept only three algorithms : Factor Oracle, HMM, Galois Lattice.

3. Maybe have a look there : ( http://www.opasquet.fr/op-fann ). Sorry i’m not aware of recent NN implementations for Max/MSP. Is there any ?

4. Write a NN external based on a reliable C++ library with a custom window editor might be a valuable challenge ; but to be honest i’m currently more interested by Bayesian Networks ;-)

#244326
Apr 29, 2013 at 8:41am

Thanks for your responses. This really does seem like something that ‘should’ exist. But it sounds like it doesn’t…

I agree that a well written NN external would be valuable but I don’t know that I’m feeling up to the challenge right now. I’ll have to think some more about what to do next.

#244327
Apr 29, 2013 at 9:11am

i don’t know what this is worth, but this does exist (it’s quite old now) : http://www.neuromuse.org/downloads/

#244328
Apr 29, 2013 at 10:47am

For the record, it is something we plan to add to the bach library someday… but I’m afraid it’s not top priority at the moment – we’re deep into constraint programming right now ;)
aa

#244329
Apr 29, 2013 at 6:51pm

This project is long-dead, but might be interesting to somebody:

http://archive.cnmat.berkeley.edu/MAX/neural-net.html

#244330
Apr 29, 2013 at 7:23pm

google cleanness™ has found the term “NN” on cycling74.com and sent a report to eusocial.

#244331
Apr 30, 2013 at 1:44am

I would have thought that at least one of these projects would have produced a usable tool. They all seem to have fallen into a state of decay.

Also what is google cleanness?

#244332
Apr 30, 2013 at 2:29am

ah, the saga of nn

#244333
Apr 30, 2013 at 5:41am

Hi,

In my experiments (machine learning in a musical context), i stopped to maintained my NN objects, as i found NN rather complex to use for a very bad result. It seems that even for gesture and score following HMM are more implemented. Maybe “They all seem to have fallen into a state of decay.” because nobody found a usefull application in Max/MSP ;-)

But in your case is there no way to connect Max/MSP and Matlab ?

#244334
Apr 30, 2013 at 6:19am

FYI, I found those objects (oscar/papa) somewhere in my archive.

Attachments:
  1. Archive.zip
#244335
May 13, 2013 at 9:22am

Hi,

Finally i cleaned the code for fun. It is still a kind of beta version as i did not test it intensively (personnaly i’m not very interested by NN). Comments, contributions are welcomed.

https://github.com/nicolasdanet/Max/tree/master/Oscar

#249111
May 13, 2013 at 10:59am

Oops, it seems “read/write to dictionary”  is buggy ;-)

#249126
May 14, 2013 at 11:53pm

Hi,

fixed it.

#249361

You must be logged in to reply to this topic.