[eu-gene] eu-gene Digest, Vol 105, Issue 35

[Mark Gould]

David, Philip,

I've been lurking on this list until now. I'm an artist and a writer quite
fascinated with generative art and related theoretical discussions like
this. I was having a similar conversation about aesthetics, computer
generated art, and research and art with colleagues recently. I wonder if
it might be possible or not, to use comments made here, with your
permission attribution, links and whatever else you might wish to be
included in a series of articles (in one case, well, blog posts)?

Thanks for your consideration.

Mark Gould


On Fri, Jan 27, 2012 at 2:58 PM, <eu-gene-request at generative.net> wrote:

> Send eu-gene mailing list submissions to
>        eu-gene at generative.net
>
> To subscribe or unsubscribe via the World Wide Web, visit
>        http://www.generative.net/mailman/listinfo/eu-gene
> or, via email, send a message with subject or body 'help' to
>        eu-gene-request at generative.net
>
> You can reach the person managing the list at
>        eu-gene-owner at generative.net
>
> When replying, please edit your Subject line so it is more specific
> than "Re: Contents of eu-gene digest..."
>
>
> Today's Topics:
>
>   1. Re: need advice on a framework design (David Hart)
>   2. Re: need advice on a framework design (Philip Galanter)
>
>
> ----------------------------------------------------------------------
>
> Message: 1
> Date: Fri, 27 Jan 2012 11:56:07 -0700
> From: David Hart <dahart at gmail.com>
> Subject: Re: [eu-gene] need advice on a framework design
> To: generative art <eu-gene at generative.net>
> Message-ID:
>        <CAFRxipHfrO6hpAnn+jdreyCBdBrKyvFv_G7YUPS8Zohq2x9OGA at mail.gmail.com
> >
> Content-Type: text/plain; charset="iso-8859-1"
>
> So, again, I do agree with some of your conclusions w.r.t. what's been done
> previously.  But rather than throw up my arms and call it a dead end, I've
> tried, in small steps, to fix the problems I see, and I suspect other
> researchers, some you've mentioned, would like to as well.  The fact that
> it only barely sorta works right now is a challenge for us, not a reason to
> give up.  The fact that it works at all only provides hope & fuel.  ;)
>
> One approach you could take is to give the user more control besides just
> selection.  If the goal is to make art, then why not?  You can guide what
> kinds of mutations take place each generation, and control the degree of
> randomness, even mix in direct manual control tools if you want.  Some
> technical folks don't like the idea of introducing other controls than
> selection, any maybe it becomes questionable whether its still evolution,
> but as a tool for producing art it accelerates evolution in the direction
> you want, thus it works well.
>
> Interactive evolution, when made more directly controllable, can be seen as
> more of an exploratory tool than as some sort of magical one-button process
> that's supposed to produce "good art" after a million generations using
> only my passive choosing of what I like or don't.  Photoshop's "variations"
> tool, for example, is a sort of small scale evolution process.  There are
> maybe five or six dimensions including hue and brightness, and I don't know
> exactly what I want so I'd like to see a set of mutations on my current
> position in this 6-D space, and I'll just choose a direction to go,
> adjusting the step size, until I get there.
>
> Genetic images, when viewed this way, can be extremely effective as a tool
> at things like texture design for 3d models.  Its a variations tool, but
> instead of only hue and brightness, we maybe start with a 300 dimensional
> space by default, and the number of parameters grows and shrinks as we go.
>  What the format of mutation & selection does really well is give us a
> framework for managing more parameters than we can possibly handle
> individually.  It simplifies exploring incredibly complicated spaces down
> to one question, 'here are some choices-- anything you see closer to what
> you want?'.  But if that's not enough control, and for making art its
> probably not enough control, start putting some control back.
>
>
> If the goal is to develop a way of extracting an aesthetic measure in an
> > automated way, then a better test bed would be manually produced art made
> > by human artists. That way you can know that whatever it is that creates
> > aesthetic value, it is in the samples in a broad range of detectable
> > concentration. Once an automated measure is in hand and verified against
> > human art, it can then be adapted for use as a fitness function.
> >
>
> I agree, and for what its worth, I believe the majority of research into
> how to systematically recognize aesthetics is already doing exactly this.
>  There are lots of people who've already tried training learning algorithms
> / fitness functions on existing databases of art.  (Also tried frequently:
> training on human faces & bodies to find "beautiful people")  I'm not sure
> anyone's produced results any more compelling than genetic images yet...
>
> Due to the conceptual & story based nature of art, this approach cannot
> achieve anything better than mediocre success at its very best unless we
> also solve artificial intelligence.  Furthermore, if it gets even that far,
> which is unlikely, then when its used as a fitness function it cannot be
> expected to produce something new, only interpolations of existing works.
>  And we're unlikely to gain understanding from this approach, a fitness
> function can't explain what makes art good it can only say yes or no.
>
>
>
> > By immediately using evolving expression images there's no guarantee that
> > aesthetic value will be in the test bed in sufficient quantity or range
> to
> > get useful results.
> >
>
> If there were guarantees, it wouldn't be evolution.  And it wouldn't be
> art.  Why do you expect some sort of "guarantee"?
>
>
> >  not to mention generative art's limited aesthetic successes, by your own
> > criteria
> >
> > I wouldn't say that at all.  Perhaps you are misapplying or
> misattributing
> > the "your own criteria" part.
> >
>
> I was referring to the criteria you laid out earlier.  "...as soon as
> evolving expressions yield a wide variety of personalized graphical styles
> I'll be happy to concede..."  What existing generative techniques currently
> meet this standard?
>
>
>
> > By comparison complexity based generative art (e.g. evolutionary art) is
> > relatively new. And I think some of it is already quite successful.
> >
>
> I'd love to see a few of your picks!
>
>
> There's aesthetic bias in every non-generative medium as well and I see no
> > reason why, from an *artistic* point of view, this is much of a
> > consideration.
> >
>
> You're introducing a different kind of bias than I was talking about.  From
> both an artistic point of view and a scientific point of view, you ought to
> be interested in making sure that you (the fitness function) are actually
> guiding the evolution rather than the evolution guiding you.
>
>
>
> > For an artist creating artifacts the evolutionary process is "working"
> > simply if it produces good art, and especially if it's good art that the
> > artist wouldn't have been likely to produce another way.
> >
>
> No, as much as you want it to be this simple, this is not a valid
> definition of a working evolutionary process.  Good art may be what the
> artists *wants*, but if he expects it then he doesn't understand his tools
> or their limits; his expectations are unrealistic and exceed the
> capabilities of evolution, artificial or real.  Artificial evolution is an
> optimization process.  The quality of its output depends on what the
> fitness function measures, and it is "working" when the output over time
> approaches greater fitness as measured by the fitness function, and nothing
> more.  If the fitness function changes every generation, like it does when
> a human pretends to be a fitness function, then all bets are off and
> getting good art out is strictly the responsibility of the artist.
>
> Also, while people are informally using genetic images as an exploratory
> tool to produce something they might not have made some other way, the
> structure of the evolution format somewhat opposes this idea.  If the
> user/artist is the fitness function, then evolution can only be expected to
> produce what they like, which if they know, then they can presumably
> produce it anyway.  Nobody claims evolution will produce results unlikely
> to be produced through some other method, and artificial evolution was not
> designed to do any such thing.  In fact it a lot cases where artificial
> evolution actually works, it has also been shown that more direct
> optimization methods, when they're available, are more efficient at finding
> the optimal solution.  Evolution is good mainly for the cases where direct
> optimization isn't available.
>
> You can assign your 'good-art' and 'unlikely-to-be-produced-another-way'
> criteria to the artist or to the art, if you want, but not to the
> evolutionary process.
>
> --
> David.
> -------------- next part --------------
> An HTML attachment was scrubbed...
> URL:
> http://www.generative.net/pipermail/eu-gene/attachments/20120127/13f0e750/attachment.html
>
> ------------------------------
>
> Message: 2
> Date: Fri, 27 Jan 2012 16:58:21 -0600
> From: Philip Galanter <list at philipgalanter.com>
> Subject: Re: [eu-gene] need advice on a framework design
> To: generative art <eu-gene at generative.net>
> Message-ID: <3B8889CC-BC11-4A94-9439-E85C633A2EE4 at philipgalanter.com>
> Content-Type: text/plain; charset="us-ascii"
>
>
> On Jan 27, 2012, at 12:56 PM, David Hart wrote:
>
> > If the goal is to develop a way of extracting an aesthetic measure in an
> automated way, then a better test bed would be manually produced art made
> by human artists. That way you can know that whatever it is that creates
> aesthetic value, it is in the samples in a broad range of detectable
> concentration. Once an automated measure is in hand and verified against
> human art, it can then be adapted for use as a fitness function.
> >
> > I agree, and for what its worth, I believe the majority of research into
> how to systematically recognize aesthetics is already doing exactly this.
>  There are lots of people who've already tried training learning algorithms
> / fitness functions on existing databases of art.
>
> Not as many as I expected to find when I looked.  Maybe I should check the
> AI literature more closely?  I'm not sure where these "lots of people" are.
>
> >  (Also tried frequently: training on human faces & bodies to find
> "beautiful people")  I'm not sure anyone's produced results any more
> compelling than genetic images yet...
>
> Again possible conflation of using genetic images as a test set to
> discover fitness functions similar to human aesthetics versus using genetic
> systems to produce aesthetically pleasing results.  Finding an evaluation
> method is in itself a valid achievement.  Applying it to create new faces
> and bodies is another step.
>
> >
> > Due to the conceptual & story based nature of art, this approach cannot
> achieve anything better than mediocre success at its very best unless we
> also solve artificial intelligence.
>
> Not necessarily. Psychologist Colin Martindale, and Daniel Berlyne before
> him, have done some interesting work suggesting neural excitation levels
> rather than specific concepts are what create a hedonic effect. Also
> preverbal children have an aesthetic response. I'd agree that higher level
> cognition makes a contribution.  But even there maybe the same basic
> neurological mechanism is involved regardless of how high or low in the
> neural hierarchy the transaction takes place.
>
> > Furthermore, if it gets even that far, which is unlikely, then when its
> used as a fitness function it cannot be expected to produce something new,
> only interpolations of existing works.
>
> But doesn't natural evolution always apply the same fitness function,
> ability to survive long enough to reproduce, and yet create new species?
>
> (This harkens back to a point I made much earlier.  If one thinks of
> evolution taking place with a single level of emergence, i.e. a direct
> mapping of genotype into phenotype, then it's hard to think of ways that
> innovation outside of interpolation can be introduced.  But if one can
> build an evolutionary system that exhibits multilevel emergence (like the
> biological world), then the complexification allowed by machines building
> machines building machines, etc can lead to true open ended innovation. My
> hunch is that evolutionary art is being held back by the lack of multilevel
> emergence as much as the "fitness bottleneck.")
>
> > And we're unlikely to gain understanding from this approach, a fitness
> function can't explain what makes art good it can only say yes or no.
> >
>
> It depends on what the fitness function is.  If someone trains up a
> complicated neural network to judge aesthetics I'd agree that those can be
> hard to tweeze apart and understand.  But in non-aesthetic evolutionary
> computing fitness functions are typically quite understandable in that the
> understanding comes first and it gets translated into a fitness function
> second.  e.g. calculating the lift of an airplane wing or the efficiency of
> a delivery route is first understood, and only then turned into a fitness
> function.  Who knows what principles psychology may reveal in the future?
>
> >
> > By immediately using evolving expression images there's no guarantee
> that aesthetic value will be in the test bed in sufficient quantity or
> range to get useful results.
> >
> > If there were guarantees, it wouldn't be evolution.  And it wouldn't be
> art.  Why do you expect some sort of "guarantee"?
>
> More confusion about what makes for useful test sets versus what is good
> for generating art.  What I'm saying is that in terms of having a test bed
> for discovering an automatic measure for aesthetics, we know without a
> doubt that human art has a wide range of aesthetic value.  We don't know a
> priori that machine generated art does. To investigate the human aesthetic
> response we need a test bed that presents the wide range of human aesthetic
> experience. So better to use human art.  We don't know that machine
> generated art presents a wide enough range to allow for the discovery of an
> automatic measure. At least not yet.
>
> >
> >>  not to mention generative art's limited aesthetic successes, by your
> own criteria
> > I wouldn't say that at all.  Perhaps you are misapplying or
> misattributing the "your own criteria" part.
> >
> > I was referring to the criteria you laid out earlier.  "...as soon as
> evolving expressions yield a wide variety of personalized graphical styles
> I'll be happy to concede..."  What existing generative techniques currently
> meet this standard?
>
> With regard to the creation of quality generative art I'd cite traditional
> generative methods using pattern, symmetry, and tiling.  They've been
> around for 1000's of years.  And then in the 20th century the application
> of chance as a generative method.
>
> But that's not the above quoted criteria.  That criteria only applies when
> a given system currently doesn't generate good results. Maybe I should
> withdraw the "wide variety" part.  I'd be happy to see one artist develop
> one personal style using evolving expressions that is different than the
> known, not terribly interesting, style.  That's what I tried to do, but
> only with marginal success.
>
> I hope someone will do better.  If you can, well great.  If someone can
> operate a spin art machine and not get paintings that look just like every
> other spin art painting, that's great too.  I'm just skeptical.  I think
> spin art will always look like spin art, and to do something different and
> better will require inventing some other kind of machine. Ditto evolving
> expressions.
>
> > There's aesthetic bias in every non-generative medium as well and I see
> no reason why, from an *artistic* point of view, this is much of a
> consideration.
> >
> > You're introducing a different kind of bias than I was talking about.
>  From both an artistic point of view and a scientific point of view, you
> ought to be interested in making sure that you (the fitness function) are
> actually guiding the evolution rather than the evolution guiding you.
>
> I'll leave the scientific point of view to one side, although it has to be
> noted that the interest is derived from the hypothesis under consideration.
> It may be different for different investigations.
>
> But from an artistic point of view, especially when the process (not the
> artifact) is indeed the art, I think an autonomous non-interactive
> evolutionary system would be *very* interesting.  Much more interesting
> than one that has to be pushed by a human to find its way.
>
>
> >
> > For an artist creating artifacts the evolutionary process is "working"
> simply if it produces good art, and especially if it's good art that the
> artist wouldn't have been likely to produce another way.
> >
> > No, as much as you want it to be this simple, this is not a valid
> definition of a working evolutionary process.
>
> I'm assuming that you can't call it an evolutionary process unless it has
> all the technical parts and functions of an evolutionary process I.e. if
> it's not an evolutionary process it shouldn't be called an evolutionary
> process.
>
> What I'm saying then is that for an artist it's not enough that the
> evolutionary process simply be technically complete.  If you had an
> evolutionary system to design airplanes, but the airplanes fell from the
> sky, I don't think you would refer to that evolutionary system as
> "working."  Similarly an artist's evolutionary system has to generate good
> art or else it's not working either...at least not in an artistic sense.
>  Perhaps a computer scientist could find love for an evolutionary system
> that designs fatal aircraft or bad art.  Not me.
> -------------- next part --------------
> An HTML attachment was scrubbed...
> URL:
> http://www.generative.net/pipermail/eu-gene/attachments/20120127/527622bd/attachment.htm
>
> ------------------------------
>
> --
> '"generative" is where you lose control of a machine which does exactly
> what you tell it'
> To unsubscribe from eu-gene visit
> http://www.generative.net/mailman/listinfo/eu-gene
>
> End of eu-gene Digest, Vol 105, Issue 35
> ****************************************
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.generative.net/pipermail/eu-gene/attachments/20120127/fc5a5d64/attachment-0001.htm