standard approach to very high dimensional optimization by genetic algorithms / machine learning methods

A place to discuss the science of computers and programs, from algorithms to computability.

Formal proofs preferred.

Moderators: phlip, Moderators General, Prelates

>-)
Posts: 510
Joined: Tue Apr 24, 2012 1:10 am UTC

standard approach to very high dimensional optimization by genetic algorithms / machine learning methods

Postby >-) » Mon Jun 13, 2016 3:32 pm UTC

Suppose you want to generate images which are pleasing to the eye (a human judge does the fitness calculations). For a 1000 x 1000 image with RGBA channels, that's 4 million inputs, which IIRC wouldn't play well with most optimization methods / neural networks and what not.

It seems that the algorithm would need to teach itself some sort of hierarchical structure in order to have any hope of producing more than white noise. For example, the highest level could deal with the theme of the image or the color palette, while the next level could deal with drawing a single object in the painting, and another level could draw a subcomponent of that object, and the bottom level could deal with dithering and aliasing. Of course, the hierarchy shouldn't need to be programmed into the algorithm -- the algorithm should figure it out on its own.

The only type of optimization technique that I'm familiar with which is capable of this is genetic programming. However, GP seems just a bit too general. Something about evolving entire computer programs feels a bit overkill to me. So my question is, is there any better way to do this / any papers / standard approaches to problems like this?

User avatar
Xanthir
My HERO!!!
Posts: 5209
Joined: Tue Feb 20, 2007 12:49 am UTC
Location: The Googleplex
Contact:

Re: standard approach to very high dimensional optimization by genetic algorithms / machine learning methods

Postby Xanthir » Mon Jun 13, 2016 8:40 pm UTC

I suggest looking into the deep learning neural nets that are popular these days; they address this precise topic in a number of ways.

At their core, they work by using multiple levels of neurons that gradually extract features from the image; starting with contrast boundaries, then edges, then shapes, etc. This turns out to be remarkably similar to how our own neural architecture processes images.
(defun fibs (n &optional (a 1) (b 1)) (take n (unfold '+ a b)))

>-)
Posts: 510
Joined: Tue Apr 24, 2012 1:10 am UTC

Re: standard approach to very high dimensional optimization by genetic algorithms / machine learning methods

Postby >-) » Tue Jun 14, 2016 3:28 am UTC

Thanks. I've seen of the examples of neural networks being used to paint one image in the style of another image, and it might be interesting to see if they can be purely creative

LjSpike
Posts: 26
Joined: Fri Jun 17, 2016 8:14 pm UTC

Re: standard approach to very high dimensional optimization by genetic algorithms / machine learning methods

Postby LjSpike » Fri Jun 24, 2016 9:49 am UTC

>-) wrote:Thanks. I've seen of the examples of neural networks being used to paint one image in the style of another image, and it might be interesting to see if they can be purely creative


Well, creating non-abstract realistic images would be tricky, but a computer could create an abstract image. It'd just need some restrictions on how small a 'block' of color is to eliminate most white noise.
Then a palette for theme (it could be given some common associations, e.g. nature = green, cold / sad = blue).

You could also give it a routine to look at the pixels of existing paintings, and then approximate 'rules' which the paintings follow, and once it has a rule created, it applies it to its own painting, thus "learning".

I can't imagine what other way you'd approach this really :/ well, without some sci-fi level AI that is.


Return to “Computer Science”

Who is online

Users browsing this forum: No registered users and 2 guests