Neural network dreams
-
- KVRist
- Topic Starter
- 225 posts since 5 Oct, 2008
Has any developer read the "Inceptionism" article from google research? Basically, once neural networks intended to do image identification are fed back on themselves, they output visual "dreams" which reflect the content with which the network has been trained. Is there any reason why this shouldn't be possible for audio? Just imagine the possibilities. Here is the article for those who might be interested: http://googleresearch.blogspot.co.uk/20 ... eural.html
-
- KVRian
- 1379 posts since 26 Apr, 2004 from UK
Yeah, typical NN things. The issue for audio is that there would be lots of noise, so maybe not that relevant.
Personal rant: 10 to 30 layers??? that's just absurd... Evenb with 2 layers we don't know what happens, so 10 to 30?
Personal rant: 10 to 30 layers??? that's just absurd... Evenb with 2 layers we don't know what happens, so 10 to 30?
-
- KVRist
- Topic Starter
- 225 posts since 5 Oct, 2008
Well, I guess feeding them midi instead of audio would generate more useful content. And I guess the more we don't know what happens, the more interesting the results are . Perhaps a layer number parameter could make things more predictable. But results would greatly depend on what you feed the network with. This would be a playground for experimental music.
- KVRAF
- 9576 posts since 16 Dec, 2002
Our ears are more choosey about what sounds right than our eyes are about interesting visuals, we can tolerate nasty visuals alot more than nasty sound
Amazon: why not use an alternative
-
- KVRist
- Topic Starter
- 225 posts since 5 Oct, 2008
Would that matter be resolved by feeding the network with musically sound material? What about note data instead of audio? If the network is trained with scores from pieces which are all in the same key, and the weighing parameters adjusted in a musical way is there a reason why it should come up with unpleasant sounding material? There's all sorts of parameters to be tweaked, and the learning algorithms are weighed, so you could favor one "dream" over another. I'm really trying to get a grip on how these things work, so forgive my ignorance.
Last edited by gustavokch on Wed Jul 01, 2015 9:16 pm, edited 1 time in total.
- KVRAF
- 9576 posts since 16 Dec, 2002
This will probably interest you
http://www.vintagesynth.com/misc/neuron.php
This is available for Mac and free
http://www.neuron-synth.com/html/Home.html
http://www.vintagesynth.com/misc/neuron.php
This is available for Mac and free
http://www.neuron-synth.com/html/Home.html
Amazon: why not use an alternative
-
- KVRist
- Topic Starter
- 225 posts since 5 Oct, 2008
Cool. I was aware of the Hartmann neuron, however processing power and neural network technology have come a long way since 2000, it's still a very interesting instrument. I wonder what sort of information is contained in it's models, and how an open design with more parameters and current day technology would fare in generating sound or musical phrases from content you feed it yourself.VariKusBrainZ wrote:This will probably interest you
http://www.vintagesynth.com/misc/neuron.php
This is available for Mac and free
http://www.neuron-synth.com/html/Home.html
-
- KVRian
- 847 posts since 20 May, 2010
If this is possible. Then this is possible:
A sound of a bird will create a picture which looks like a bird.
A picture or gif animation of a bird will then create a sound audio file which sounds like a bird.
The neural network will approximate how the bird photo will sound like as a audio file based on nature videos it has seen earlier.
The neural network will approximate how the bird sound recording looks like as a picture or gif animation based on nature videos it has seen earlier.
Humans can do this with their brain imagination by remembering which sounds and looks like what.
What the eyes can see that the ears can't hear will produce unpredictable sounds.
What the ears can hear that the eyes can't see will produce unpredictable pictures and gif animations.
A sound of a bird will create a picture which looks like a bird.
A picture or gif animation of a bird will then create a sound audio file which sounds like a bird.
The neural network will approximate how the bird photo will sound like as a audio file based on nature videos it has seen earlier.
The neural network will approximate how the bird sound recording looks like as a picture or gif animation based on nature videos it has seen earlier.
Humans can do this with their brain imagination by remembering which sounds and looks like what.
What the eyes can see that the ears can't hear will produce unpredictable sounds.
What the ears can hear that the eyes can't see will produce unpredictable pictures and gif animations.