[dorkbotdc-blabber] some blather about last nights dork

[Nikolas Coukouma]

Philip Kohn wrote:
> Pindar's talk really got me thinking about creativity.
> I would really like to know more about how things like balance, and
> composition
> are handled without randomness.
> I really love the idea of seeing how far you can get without any
> random numbers.
> I've always relied on some sort of random functions, even if they are
> carefully tuned and filtered,
> but in a way it is really a cop out.
I'm interested in discussing this more ... possibly in an in-person
meeting/get-together. I hope to stop by HacDC tonight, so if you (or
other interested people) are there, that's all the merrier.

A bit of background: I am a computer scientist; some would call me a
mathematician as well, but I lack training in that field. I tend to take
an extremely technical approach to things, which I recognize as both
useful and limiting ;) I have some creative hobbies (primarily
photography) and a general interest in "creative stuff."

Back to the subject at-hand ...
In my opinion, you can not have a system that is (independently)
creative without randomness of some sort. My reasoning is that, without
randomness, you are limited to a strict set of rules, which were
designed by some human. I'm guessing that this objection of randomness
is "use a random number as an input"; it could easily result just from
measurement errors and the like ... the uncertainty inherent in the world.

> Most of the tools of artificial intelligence (neural nets, backprop,
> etc.) are designed to capture
> regularities in the input/output transformation.
Yes.
Below is a lengthy digression to define some terms...
* Artificial intelligence deals with "intelligent agents" - systems that
perceive the state of their world and then take an action to accomplish
a goal (whether its moving to a different location or changing an answer
on a multiple choice test). Usually some metric is defined so its
success can be measured.
* Artificial neural networks consist of a connected system of
nodes/vertexes, inspired by biological neurons (and bear about as much
resemblance as many novels "inspired by a true story").  Each node takes
one or more input and produces one output. That output can then be sent
to one or more other neurons. There is no randomness; these are
functions in the mathematical sense.
* Backprogation is a "supervised" method for training neural networks.
The network is given a test input, its output is compared to the
expected output. Then blame is assigned for errors and adjustment is
performed. The adjustments start closest to the output and proceed back
towards the input. It's supervised because, although you're not
necessarily sitting there, you have decided what the right answer is.
* K-means clustering takes a collection of n objects (e.g. DVDs) with
multiple attributes (e.g. each users rating is a different attribute)
and puts them into k groups (e.g. 10, for whatever reason you picked
it). Each item can only be assigned to one group. The goal is to make
each group as similar as possible.
* Genetic algorithms search for an optimal solution by generating a
set/population of possible solutions and then "breeding" them. The
breeding part consists of determining how good each solution is,
deciding whether or not it should be discarded, mating the survivors
("crossover") to produce new solutions, and randomly each of these new
solutions a bit. The crossover step keeps different parts of each
mate... the core of genetic algorithms has a mathematical basis and,
again, is only "inspired by" biology.

> But I think art and creativity revolve around the interplay of
> expectations and surprises.
> You have to have rules, but you also have to break them, and break
> them the right amount and
> in the right contexts.
> You need AI tools that can take a model of the regularities and figure
> out how to make these interesting
> and strong exceptions.
> The worst thing you can do as an artist is to create something that
> looks like a mistake.
> (Although you can repeat that "mistake" and then it may become good
> art again!)
> The exceptions need to stand out.
See genetic algorithms above; there's also a vast collection of
statistical approaches, which are particularly popular for natural
language stuff. A fun example is this blog haiku generator
http://memes.angrygoats.net/forms/haiku
http://memes.angrygoats.net/faq

An example from http://memes.angrygoats.net/livejournal.com/nofcna/haiku

transforming something
intangible gives meaning
to its existence

> I have done some art evolution, using my own ratings as a fitness
> function.
This seems like a reasonable approach to me
> Here are some of the best links:
Thanks for these

> There is a big problem with all this.
> When you have a lot of parameters, or knobs to twiddle, you need a
> minimum of twice that number of example datasets.
> So if you really want to evolve each brush stroke, you will need a lot
> of user input.

There's quite a bit of interest in this ... among other things, Amazon
has a service for paying people to complete these sorts of tasks:
http://www.mturk.com/mturk/welcome
A non-monetary incentive for people to participate is fun. I know there
are at least a few games online there that are being used for data
collection...
http://www.gwap.com/
Rating communities (e.g. Flickr, Amazon) are another possible source of
information... some of the work on incentives for distributed computing
(donating your computer's time to some task) may also be worth looking
into (e.g. lottrees).

>
> One solution would be to evolve a "critic" that evaluates the results,
> and then let it work automatically.
> I have some ideas about how to evolve the critic based on examples of
> good art that can be found easily on the web. 
A related concept in genetic algorithms is "co-evolution"; in chess you
might have a collection of black players and a collection of white
players, and have them play against each other as they evolve.

Cheers,
-Nikolas

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 189 bytes
Desc: OpenPGP digital signature
Url : http://music.columbia.edu/pipermail/dorkbotdc-blabber/attachments/20080528/178ef901/signature.bin