Loading ...
Sorry, an error occurred while loading the content.

Re: [neat] Re: Picbreeder comments

Expand Messages
  • John Arrowwood
    ... In classical art there are rules for good works. The rule of thirds, for example. Why do those rules exist? I would posit that those rules exist
    Message 1 of 17 , Sep 4, 2007
    • 0 Attachment
      On 8/30/07, Kenneth Stanley <kstanley@...> wrote:
      > I agree with Peter and Jeff that it would be very difficult to learn
      > what is interesting, even from a relatively good data set (like
      > Picbreeder perhaps could provide). The problem is that there are so
      > many possible spurious correlations between good pictures and other
      > irrelevent factors that the classifier will not get the deeper
      > concept of what makes a good picture.
      >
      > Yet I agree with John in the sense that there is extremely rich and
      > unusual data in the Picbreeder database that may be useful for
      > something beyond Picbreeder itself. I just think the right way to
      > leverage it is tricky.
      >
      > The idea of complexity is also quite muddy. Unfortunately, the word
      > means a lot of different things. In fact, perhaps the best formal
      > definition ("minimal description length") is both proven
      > uncomputable in general and also does not correlate to what is meant
      > by the normal intuitive use of the word complexity with respect to
      > evolution.
      >
      > As evolutionary complexity goes up in the sense we usually mean it,
      > it is not just that the phenotype takes more bits to describe.
      > There is also a sense of organization, i.e. regularity, that begins
      > to take hold. That is in direct opposition with the idea of minimal
      > description length. In fact, what we really want is some kind of
      > partnership between complexity and organization. Unfortunately that
      > introduces yet another mysterious word: organization; it also
      > introduced a partnership between two seemingly opposing forces. Yet
      > I think this muddiness and internal conflict is important because it
      > shows that this problem is extremely slippery. It makes me wonder
      > what is meant by "doing well" for humans asked to select for
      > complexity.
      >
      > In any case, I do like the theme behind John's idea, but I think
      > something more sophisitcated needs to be done to capture the deeper
      > reasons that certain pictures are selected.

      In classical art there are 'rules' for good works. The rule of
      thirds, for example. Why do those rules exist? I would posit that
      those rules exist because of a mathematical relationship that exists
      in how the signals travel around in our brain. I suspect that a
      network COULD be developed that would be able to identify the pure
      aesthetic merit of a particular image, because I think that at that
      level, aesthetics is a purely mathematical function.

      That said, I think that it would be very unlikely to be able to say,
      "hey, that looks like a bunny rabbit that has been borgified...cool!"
      Since our choice of what to breed for is based partly on aesthetics
      and partly on the emotional tie to the objects that something reminds
      us of, there is little chance that a controller network could totally
      replace the human.

      But what if? What if a controller could be evolved which learned to
      recognize good image composition rules? How might such a thing extend
      the usefulness of PicBreeder in general? What might it teach us?
      What might it say about the limits of computer controls? What might
      it say about US? Is there any merit at all to undertaking such an
      experiment? In my opinion, there is.

      -- John
      --
      John Arrowwood
      John (at) Hanlons Razor (dot) com
      John (at) Arrowwood Photography (dot) com
      --
      http://arrowwood.blogspot.com
    • Kenneth Stanley
      ... learn ... are so ... other ... and ... to ... word ... formal ... meant ... to ... it, ... begins ... minimal ... that ... Yet ... because it ... wonder
      Message 2 of 17 , Sep 9, 2007
      • 0 Attachment
        --- In neat@yahoogroups.com, "John Arrowwood" <jarrowwx@...> wrote:
        >
        > On 8/30/07, Kenneth Stanley <kstanley@...> wrote:
        > > I agree with Peter and Jeff that it would be very difficult to
        learn
        > > what is interesting, even from a relatively good data set (like
        > > Picbreeder perhaps could provide). The problem is that there
        are so
        > > many possible spurious correlations between good pictures and
        other
        > > irrelevent factors that the classifier will not get the deeper
        > > concept of what makes a good picture.
        > >
        > > Yet I agree with John in the sense that there is extremely rich
        and
        > > unusual data in the Picbreeder database that may be useful for
        > > something beyond Picbreeder itself. I just think the right way
        to
        > > leverage it is tricky.
        > >
        > > The idea of complexity is also quite muddy. Unfortunately, the
        word
        > > means a lot of different things. In fact, perhaps the best
        formal
        > > definition ("minimal description length") is both proven
        > > uncomputable in general and also does not correlate to what is
        meant
        > > by the normal intuitive use of the word complexity with respect
        to
        > > evolution.
        > >
        > > As evolutionary complexity goes up in the sense we usually mean
        it,
        > > it is not just that the phenotype takes more bits to describe.
        > > There is also a sense of organization, i.e. regularity, that
        begins
        > > to take hold. That is in direct opposition with the idea of
        minimal
        > > description length. In fact, what we really want is some kind of
        > > partnership between complexity and organization. Unfortunately
        that
        > > introduces yet another mysterious word: organization; it also
        > > introduced a partnership between two seemingly opposing forces.
        Yet
        > > I think this muddiness and internal conflict is important
        because it
        > > shows that this problem is extremely slippery. It makes me
        wonder
        > > what is meant by "doing well" for humans asked to select for
        > > complexity.
        > >
        > > In any case, I do like the theme behind John's idea, but I think
        > > something more sophisitcated needs to be done to capture the
        deeper
        > > reasons that certain pictures are selected.
        >
        > In classical art there are 'rules' for good works. The rule of
        > thirds, for example. Why do those rules exist? I would posit that
        > those rules exist because of a mathematical relationship that exists
        > in how the signals travel around in our brain. I suspect that a
        > network COULD be developed that would be able to identify the pure
        > aesthetic merit of a particular image, because I think that at that
        > level, aesthetics is a purely mathematical function.
        >
        > That said, I think that it would be very unlikely to be able to say,
        > "hey, that looks like a bunny rabbit that has been
        borgified...cool!"
        > Since our choice of what to breed for is based partly on aesthetics
        > and partly on the emotional tie to the objects that something
        reminds
        > us of, there is little chance that a controller network could
        totally
        > replace the human.
        >
        > But what if? What if a controller could be evolved which learned to
        > recognize good image composition rules? How might such a thing
        extend
        > the usefulness of PicBreeder in general? What might it teach us?
        > What might it say about the limits of computer controls? What might
        > it say about US? Is there any merit at all to undertaking such an
        > experiment? In my opinion, there is.
        >
        > -- John

        John, you argue persuasively for your idea, and I can see the
        appeal. I would not dismiss the idea that there is the possibility
        that a classifier could evolve that detects something surprising
        about aesthetic value and thereby filters the search for interesting
        images. It is an intoxicating idea and I would love to hear about
        and see the results of any such attempt.

        Yet let me give you an idea about the origin of my skepticism. I'm
        basically jaded by the underwhelming results of past such efforts.
        For example, a friend of mine tried to evolve a neural network to
        detect good music in a similar approach to yours. I have heard of
        other attempts like that one (usually they seem to be with music). I
        realize these stories are only anecdotal, yet I draw a lesson from
        them which shows the danger in such an approach:

        Evolution likes to cheat and get by easy. If you want it to learn
        something deep and general, it tries to learn some cheap trick to do
        almost as well without having to deal with the deep stuff. In other
        words, like lightning, it takes the path of least resistance. The
        problem with training to detect artistic quality is that there may be
        predominant correlations among pictures or songs that we deem good
        that are really not the reason they are good.

        For exampele, "good" songs may tend to have a lot of C notes, not
        because C makes a good song, but rather because it happens to appear
        a lot in a particular genre, perhaps because it is the base note of
        some scale. Thus, an evolved classifier will simply predict that
        anything with a lot of C's is good, missing the deeper reasons
        entirely. Or, in Picbreeder, you may notice if you look closely that
        many (though not all) of the published images have some sort of small
        circular formation near their center. For example this object may be
        a nose, an eye (in a profile), a cockpit window, an iris, a keyhole,
        a sun, a moon, a fingernail, etc. Therefore, a classifier would do
        well to learn to predict that any picture with a small circle in the
        middle is good, again missing the point.

        I think the reason such projects have failed in the past is because
        of just such spurious correlations. The question is how to push
        evolution to learn the deeper meaning of "good," which is so
        profoundly slippery.

        Nevertheless, I would not say it is a lost cause. Yet it is
        certainly a grand challenge and will not succumb easily.

        ken
      • John Arrowwood
        ... [snip] I have to agree, you are right. The tendency to take a short-cut can be a problem. That is why I was so excited by the study I saw done where a
        Message 3 of 17 , Sep 9, 2007
        • 0 Attachment
          On 9/9/07, Kenneth Stanley <kstanley@...> wrote:
          > John, you argue persuasively for your idea, and I can see the
          > appeal. I would not dismiss the idea that there is the possibility
          > that a classifier could evolve that detects something surprising
          > about aesthetic value and thereby filters the search for interesting
          > images. It is an intoxicating idea and I would love to hear about
          > and see the results of any such attempt.
          >
          > Yet let me give you an idea about the origin of my skepticism.
          [snip]

          I have to agree, you are right. The tendency to take a short-cut can
          be a problem. That is why I was so excited by the study I saw done
          where a network was explicitly punished for doing so. The results in
          that case were nothing short of stunning.

          But I wasn't proposing to evolve the controller using published images
          from PicBreeder. I was proposing that it be trained to judge between
          crops of real images, crops that have aesthetically pleasing
          qualities, and randomly generated images of various kinds. That
          should be a fairly simple, straight forward exercise. It should learn
          to detect 'signal' vs. noise, a simple mathematical function, right?
          As it plateaus at that task, start throwing more and more 'orderly'
          noise at it, and/or noise of different kinds. That is, GUIDE the
          evolution. As it learns to tell the difference between 'normal' and
          'noise' images, start varying the parameters of the 'noise.' Over
          time, if the path that the 'noise' images takes is controlled, you
          should be able to lead the network TOWARDS the 'deeper' function that
          it is meant to learn. Maybe something like this:

          * First learn to distinguish between image and a uniform field of one
          color. It'll probably cheat on this task, but that's okay.
          * Then between image and nearly uniform, with, say, 1% pixel noise.
          * As it gets better at dealing with the pixel noise, increase the
          pixel noise levels
          * When it can deal with large pixel noise, start varying the size of
          the pixels. Add the 'noise' by adding 'circles' from 1-n pixels in
          diameter. As the value of n increases, the 'orderliness' of the
          'random' image increases, making it harder and harder to tell the
          difference based on a 'short-cut' and forcing a successful network to
          evolve a more intelligent solution, though not necessarily a solution
          that gives us what we want, just yet.
          * When it can do that, start mixing circles of various sizes with
          lines of various thicknesses and lengths.
          * As we do this, we continue to include the solid-color and various
          levels of pixel noise test samples in the test, so that the network is
          expected to remember how to do what it has done before
          * When the noise reaches a certain level, we start taking the 'nice'
          pictures, and applying noise to them. The successful network will
          have learned to tell what is noise and what is not, and will be able
          to rate the non-noisy version of an image higher than the version that
          has had noise added to it.

          Up to this point, the jumps are all pretty easy, and it should be
          doable. Making the final jump is harder. We don't KNOW necessarily
          what makes for an 'aesthetically pleasing' image, we just know that it
          must have some kind of 'structure' to it, which we have been teaching
          the network to find.

          But it is not just structure, it must have some certain
          characteristics to that structure in order to be pleasing. There must
          be a counter-example to learn to distinguish between. This is where
          the process starts to get difficult...

          And it's late. If I come up with an elegant way to make that jump,
          I'll let you know! :)

          --
          John Arrowwood
          John (at) Hanlons Razor (dot) com
          John (at) Arrowwood Photography (dot) com
          --
          http://arrowwood.blogspot.com
        • kelly parker
          Excuse me while I ramble. I think what you are attempting to evolve is marketability not art. True art has no mass. The physical embodiment does. Many people
          Message 4 of 17 , Sep 10, 2007
          • 0 Attachment
            Excuse me while I ramble.

            I think what you are attempting to evolve is marketability not art.

            True art has no mass.
            The physical embodiment does.

            Many people mistake skill for talent.
            This is easy enough to understand considering that cow manure can be marketed as fertilizer if properly packaged.
            They are not the same thing.

            The rules for art are simple
            1) There are no rules.
            2) Null

            The rules for skill are also simple.
            Duplicate accepted behavior as closely as possible.

            The rules for marketability of "art".
            Herd familiarity with the basic structure.
            mild/mediocre/average/non offensive

            ref: American Idol (high skill, high marketability, little to no talent....but a cute tush) The winners invariably sound like someone familiar because they have the skill to reproduce the style of a previously popular musician.

            A CD can duplicate the greatest music ever composed.
            A camera can duplicate the greatest paintings.
            A photocopier can duplicate the greatest poetry.
            A CNC router can duplicate the greatest sculpture.
            You could stretch this to say they (the machines) have skill but I don't think anybody would mistake this for talent.

            If you study artists instead of art you'll find that a lot of them have/had "mental illnesses". (that is the artists head doesn't work like the herd thinks it should)
            Many people would argue that that "illness" is exactly what made them great.

            So with all that mess out of the way...

            I can't imagine any way to code talent but an interesting experiment might be to use lots of the "magic numbers" (Pi, Fibonacci, etc.) in an algo. and then try to figure out how to make it "crazy".

            Skill and marketability on the other hand should be trivial.
            Skill exists by default.
            marketability would be simply morphing graphical works that have already proven popular.

            for what it's worth,
            K







            Kenneth Stanley <kstanley@...> wrote:
            --- In neat@yahoogroups. com, "John Arrowwood" <jarrowwx@.. .> wrote:
            >
            > On 8/30/07, Kenneth Stanley <kstanley@.. .> wrote:
            > > I agree with Peter and Jeff that it would be very difficult to
            learn
            > > what is interesting, even from a relatively good data set (like
            > > Picbreeder perhaps could provide). The problem is that there
            are so
            > > many possible spurious correlations between good pictures and
            other
            > > irrelevent factors that the classifier will not get the deeper
            > > concept of what makes a good picture.
            > >
            > > Yet I agree with John in the sense that there is extremely rich
            and
            > > unusual data in the Picbreeder database that may be useful for
            > > something beyond Picbreeder itself. I just think the right way
            to
            > > leverage it is tricky.
            > >
            > > The idea of complexity is also quite muddy. Unfortunately, the
            word
            > > means a lot of different things. In fact, perhaps the best
            formal
            > > definition ("minimal description length") is both proven
            > > uncomputable in general and also does not correlate to what is
            meant
            > > by the normal intuitive use of the word complexity with respect
            to
            > > evolution.
            > >
            > > As evolutionary complexity goes up in the sense we usually mean
            it,
            > > it is not just that the phenotype takes more bits to describe.
            > > There is also a sense of organization, i.e. regularity, that
            begins
            > > to take hold. That is in direct opposition with the idea of
            minimal
            > > description length. In fact, what we really want is some kind of
            > > partnership between complexity and organization. Unfortunately
            that
            > > introduces yet another mysterious word: organization; it also
            > > introduced a partnership between two seemingly opposing forces.
            Yet
            > > I think this muddiness and internal conflict is important
            because it
            > > shows that this problem is extremely slippery. It makes me
            wonder
            > > what is meant by "doing well" for humans asked to select for
            > > complexity.
            > >
            > > In any case, I do like the theme behind John's idea, but I think
            > > something more sophisitcated needs to be done to capture the
            deeper
            > > reasons that certain pictures are selected.
            >
            > In classical art there are 'rules' for good works. The rule of
            > thirds, for example. Why do those rules exist? I would posit that
            > those rules exist because of a mathematical relationship that exists
            > in how the signals travel around in our brain. I suspect that a
            > network COULD be developed that would be able to identify the pure
            > aesthetic merit of a particular image, because I think that at that
            > level, aesthetics is a purely mathematical function.
            >
            > That said, I think that it would be very unlikely to be able to say,
            > "hey, that looks like a bunny rabbit that has been
            borgified... cool!"
            > Since our choice of what to breed for is based partly on aesthetics
            > and partly on the emotional tie to the objects that something
            reminds
            > us of, there is little chance that a controller network could
            totally
            > replace the human.
            >
            > But what if? What if a controller could be evolved which learned to
            > recognize good image composition rules? How might such a thing
            extend
            > the usefulness of PicBreeder in general? What might it teach us?
            > What might it say about the limits of computer controls? What might
            > it say about US? Is there any merit at all to undertaking such an
            > experiment? In my opinion, there is.
            >
            > -- John

            John, you argue persuasively for your idea, and I can see the
            appeal. I would not dismiss the idea that there is the possibility
            that a classifier could evolve that detects something surprising
            about aesthetic value and thereby filters the search for interesting
            images. It is an intoxicating idea and I would love to hear about
            and see the results of any such attempt.

            Yet let me give you an idea about the origin of my skepticism. I'm
            basically jaded by the underwhelming results of past such efforts.
            For example, a friend of mine tried to evolve a neural network to
            detect good music in a similar approach to yours. I have heard of
            other attempts like that one (usually they seem to be with music). I
            realize these stories are only anecdotal, yet I draw a lesson from
            them which shows the danger in such an approach:

            Evolution likes to cheat and get by easy. If you want it to learn
            something deep and general, it tries to learn some cheap trick to do
            almost as well without having to deal with the deep stuff. In other
            words, like lightning, it takes the path of least resistance. The
            problem with training to detect artistic quality is that there may be
            predominant correlations among pictures or songs that we deem good
            that are really not the reason they are good.

            For exampele, "good" songs may tend to have a lot of C notes, not
            because C makes a good song, but rather because it happens to appear
            a lot in a particular genre, perhaps because it is the base note of
            some scale. Thus, an evolved classifier will simply predict that
            anything with a lot of C's is good, missing the deeper reasons
            entirely. Or, in Picbreeder, you may notice if you look closely that
            many (though not all) of the published images have some sort of small
            circular formation near their center. For example this object may be
            a nose, an eye (in a profile), a cockpit window, an iris, a keyhole,
            a sun, a moon, a fingernail, etc. Therefore, a classifier would do
            well to learn to predict that any picture with a small circle in the
            middle is good, again missing the point.

            I think the reason such projects have failed in the past is because
            of just such spurious correlations. The question is how to push
            evolution to learn the deeper meaning of "good," which is so
            profoundly slippery.

            Nevertheless, I would not say it is a lost cause. Yet it is
            certainly a grand challenge and will not succumb easily.

            ken



            Moody friends. Drama queens. Your life? Nope! - their life, your story.
            Play Sims Stories at Yahoo! Games.

          • Kenneth Stanley
            John, I see where you re going. I agree that this kind of guided process will get us closer to the ideal filter. Of course, the question is how close. Yet
            Message 5 of 17 , Sep 16, 2007
            • 0 Attachment
              John, I see where you're going. I agree that this kind of guided
              process will get us closer to the ideal filter. Of course, the
              question is how close. Yet even modest filtering could be
              interesting and useful, so it might be nice to have such a thing.
              It's probably worth a shot to see if anything useful comes out of it.

              By the way, could you remind me what that study was that
              was "nothing short of stunning?" I can't remember if you cited it
              already, but I wonder if you have more detail on that? It
              definitely sounds interesting.

              ken

              --- In neat@yahoogroups.com, "John Arrowwood" <jarrowwx@...> wrote:
              >
              > On 9/9/07, Kenneth Stanley <kstanley@...> wrote:
              > > John, you argue persuasively for your idea, and I can see the
              > > appeal. I would not dismiss the idea that there is the
              possibility
              > > that a classifier could evolve that detects something surprising
              > > about aesthetic value and thereby filters the search for
              interesting
              > > images. It is an intoxicating idea and I would love to hear
              about
              > > and see the results of any such attempt.
              > >
              > > Yet let me give you an idea about the origin of my skepticism.
              > [snip]
              >
              > I have to agree, you are right. The tendency to take a short-cut
              can
              > be a problem. That is why I was so excited by the study I saw done
              > where a network was explicitly punished for doing so. The results
              in
              > that case were nothing short of stunning.
              >
              > But I wasn't proposing to evolve the controller using published
              images
              > from PicBreeder. I was proposing that it be trained to judge
              between
              > crops of real images, crops that have aesthetically pleasing
              > qualities, and randomly generated images of various kinds. That
              > should be a fairly simple, straight forward exercise. It should
              learn
              > to detect 'signal' vs. noise, a simple mathematical function,
              right?
              > As it plateaus at that task, start throwing more and more 'orderly'
              > noise at it, and/or noise of different kinds. That is, GUIDE the
              > evolution. As it learns to tell the difference between 'normal'
              and
              > 'noise' images, start varying the parameters of the 'noise.' Over
              > time, if the path that the 'noise' images takes is controlled, you
              > should be able to lead the network TOWARDS the 'deeper' function
              that
              > it is meant to learn. Maybe something like this:
              >
              > * First learn to distinguish between image and a uniform field of
              one
              > color. It'll probably cheat on this task, but that's okay.
              > * Then between image and nearly uniform, with, say, 1% pixel noise.
              > * As it gets better at dealing with the pixel noise, increase the
              > pixel noise levels
              > * When it can deal with large pixel noise, start varying the size
              of
              > the pixels. Add the 'noise' by adding 'circles' from 1-n pixels in
              > diameter. As the value of n increases, the 'orderliness' of the
              > 'random' image increases, making it harder and harder to tell the
              > difference based on a 'short-cut' and forcing a successful network
              to
              > evolve a more intelligent solution, though not necessarily a
              solution
              > that gives us what we want, just yet.
              > * When it can do that, start mixing circles of various sizes with
              > lines of various thicknesses and lengths.
              > * As we do this, we continue to include the solid-color and various
              > levels of pixel noise test samples in the test, so that the
              network is
              > expected to remember how to do what it has done before
              > * When the noise reaches a certain level, we start taking
              the 'nice'
              > pictures, and applying noise to them. The successful network will
              > have learned to tell what is noise and what is not, and will be
              able
              > to rate the non-noisy version of an image higher than the version
              that
              > has had noise added to it.
              >
              > Up to this point, the jumps are all pretty easy, and it should be
              > doable. Making the final jump is harder. We don't KNOW
              necessarily
              > what makes for an 'aesthetically pleasing' image, we just know
              that it
              > must have some kind of 'structure' to it, which we have been
              teaching
              > the network to find.
              >
              > But it is not just structure, it must have some certain
              > characteristics to that structure in order to be pleasing. There
              must
              > be a counter-example to learn to distinguish between. This is
              where
              > the process starts to get difficult...
              >
              > And it's late. If I come up with an elegant way to make that jump,
              > I'll let you know! :)
              >
              > --
              > John Arrowwood
              > John (at) Hanlons Razor (dot) com
              > John (at) Arrowwood Photography (dot) com
              > --
              > http://arrowwood.blogspot.com
              >
            • Kenneth Stanley
              The problem is that it s easy to get crazy just by outputting a bunch of random numbers; yet that is not what artists generally seem to be doing. I think
              Message 6 of 17 , Sep 16, 2007
              • 0 Attachment
                The problem is that it's easy to get "crazy" just by outputting a
                bunch of random numbers; yet that is not what artists generally seem
                to be doing. I think what art does is show you something familiar
                in a way you haven't seen it before. So it's kind of a variation on
                a theme, or a new angle. It doesn't look that far outside the realm
                of possibility that a computer could do that, although it is a very
                tough problem.

                I think one of the biggest difficulties with art is its emotional
                content. While a computer may present something in a surprisingly
                novel way, it is always somehow devoid of emotional significance,
                since the computer is not coming from a subjective perspective the
                way an artist would be. The shading in a scene may evoke the
                sadness the artist remembers on a particular lonely night. Yet for
                a computer it is merely an arbitrary filter on the colors. The time
                when computers really infuse their works with emotional content is
                certainly far away.

                ken

                --- In neat@yahoogroups.com, kelly parker <bigblockfw@...> wrote:
                >
                > Excuse me while I ramble.
                >
                > I think what you are attempting to evolve is marketability not art.
                >
                > True art has no mass.
                > The physical embodiment does.
                >
                > Many people mistake skill for talent.
                > This is easy enough to understand considering that cow manure can
                be marketed as fertilizer if properly packaged.
                > They are not the same thing.
                >
                > The rules for art are simple
                > 1) There are no rules.
                > 2) Null
                >
                > The rules for skill are also simple.
                > Duplicate accepted behavior as closely as possible.
                >
                > The rules for marketability of "art".
                > Herd familiarity with the basic structure.
                > mild/mediocre/average/non offensive
                >
                > ref: American Idol (high skill, high marketability, little to no
                talent....but a cute tush) The winners invariably sound like someone
                familiar because they have the skill to reproduce the style of a
                previously popular musician.
                >
                > A CD can duplicate the greatest music ever composed.
                > A camera can duplicate the greatest paintings.
                > A photocopier can duplicate the greatest poetry.
                > A CNC router can duplicate the greatest sculpture.
                > You could stretch this to say they (the machines) have skill but I
                don't think anybody would mistake this for talent.
                >
                > If you study artists instead of art you'll find that a lot of them
                have/had "mental illnesses". (that is the artists head doesn't work
                like the herd thinks it should)
                > Many people would argue that that "illness" is exactly what made
                them great.
                >
                > So with all that mess out of the way...
                >
                > I can't imagine any way to code talent but an interesting
                experiment might be to use lots of the "magic numbers" (Pi,
                Fibonacci, etc.) in an algo. and then try to figure out how to make
                it "crazy".
                >
                > Skill and marketability on the other hand should be trivial.
                > Skill exists by default.
                > marketability would be simply morphing graphical works that have
                already proven popular.
                >
                > for what it's worth,
                > K
                >
                >
                >
                >
                >
                >
                >
                > Kenneth Stanley <kstanley@...>
                wrote: --- In
                neat@yahoogroups.com, "John Arrowwood" <jarrowwx@> wrote:
                > >
                > > On 8/30/07, Kenneth Stanley <kstanley@> wrote:
                > > > I agree with Peter and Jeff that it would be very difficult
                to
                > learn
                > > > what is interesting, even from a relatively good data set
                (like
                > > > Picbreeder perhaps could provide). The problem is that
                there
                > are so
                > > > many possible spurious correlations between good pictures
                and
                > other
                > > > irrelevent factors that the classifier will not get the
                deeper
                > > > concept of what makes a good picture.
                > > >
                > > > Yet I agree with John in the sense that there is extremely
                rich
                > and
                > > > unusual data in the Picbreeder database that may be useful
                for
                > > > something beyond Picbreeder itself. I just think the right
                way
                > to
                > > > leverage it is tricky.
                > > >
                > > > The idea of complexity is also quite muddy. Unfortunately,
                the
                > word
                > > > means a lot of different things. In fact, perhaps the best
                > formal
                > > > definition ("minimal description length") is both proven
                > > > uncomputable in general and also does not correlate to what
                is
                > meant
                > > > by the normal intuitive use of the word complexity with
                respect
                > to
                > > > evolution.
                > > >
                > > > As evolutionary complexity goes up in the sense we usually
                mean
                > it,
                > > > it is not just that the phenotype takes more bits to
                describe.
                > > > There is also a sense of organization, i.e. regularity, that
                > begins
                > > > to take hold. That is in direct opposition with the idea of
                > minimal
                > > > description length. In fact, what we really want is some
                kind of
                > > > partnership between complexity and organization.
                Unfortunately
                > that
                > > > introduces yet another mysterious word: organization; it also
                > > > introduced a partnership between two seemingly opposing
                forces.
                > Yet
                > > > I think this muddiness and internal conflict is important
                > because it
                > > > shows that this problem is extremely slippery. It makes me
                > wonder
                > > > what is meant by "doing well" for humans asked to select for
                > > > complexity.
                > > >
                > > > In any case, I do like the theme behind John's idea, but I
                think
                > > > something more sophisitcated needs to be done to capture the
                > deeper
                > > > reasons that certain pictures are selected.
                > >
                > > In classical art there are 'rules' for good works. The rule of
                > > thirds, for example. Why do those rules exist? I would posit
                that
                > > those rules exist because of a mathematical relationship that
                exists
                > > in how the signals travel around in our brain. I suspect that a
                > > network COULD be developed that would be able to identify the
                pure
                > > aesthetic merit of a particular image, because I think that at
                that
                > > level, aesthetics is a purely mathematical function.
                > >
                > > That said, I think that it would be very unlikely to be able to
                say,
                > > "hey, that looks like a bunny rabbit that has been
                > borgified...cool!"
                > > Since our choice of what to breed for is based partly on
                aesthetics
                > > and partly on the emotional tie to the objects that something
                > reminds
                > > us of, there is little chance that a controller network could
                > totally
                > > replace the human.
                > >
                > > But what if? What if a controller could be evolved which
                learned to
                > > recognize good image composition rules? How might such a thing
                > extend
                > > the usefulness of PicBreeder in general? What might it teach
                us?
                > > What might it say about the limits of computer controls? What
                might
                > > it say about US? Is there any merit at all to undertaking such
                an
                > > experiment? In my opinion, there is.
                > >
                > > -- John
                >
                > John, you argue persuasively for your idea, and I can see the
                > appeal. I would not dismiss the idea that there is the
                possibility
                > that a classifier could evolve that detects something surprising
                > about aesthetic value and thereby filters the search for
                interesting
                > images. It is an intoxicating idea and I would love to hear
                about
                > and see the results of any such attempt.
                >
                > Yet let me give you an idea about the origin of my skepticism.
                I'm
                > basically jaded by the underwhelming results of past such
                efforts.
                > For example, a friend of mine tried to evolve a neural network to
                > detect good music in a similar approach to yours. I have heard
                of
                > other attempts like that one (usually they seem to be with
                music). I
                > realize these stories are only anecdotal, yet I draw a lesson
                from
                > them which shows the danger in such an approach:
                >
                > Evolution likes to cheat and get by easy. If you want it to
                learn
                > something deep and general, it tries to learn some cheap trick to
                do
                > almost as well without having to deal with the deep stuff. In
                other
                > words, like lightning, it takes the path of least resistance.
                The
                > problem with training to detect artistic quality is that there
                may be
                > predominant correlations among pictures or songs that we deem
                good
                > that are really not the reason they are good.
                >
                > For exampele, "good" songs may tend to have a lot of C notes, not
                > because C makes a good song, but rather because it happens to
                appear
                > a lot in a particular genre, perhaps because it is the base note
                of
                > some scale. Thus, an evolved classifier will simply predict that
                > anything with a lot of C's is good, missing the deeper reasons
                > entirely. Or, in Picbreeder, you may notice if you look closely
                that
                > many (though not all) of the published images have some sort of
                small
                > circular formation near their center. For example this object
                may be
                > a nose, an eye (in a profile), a cockpit window, an iris, a
                keyhole,
                > a sun, a moon, a fingernail, etc. Therefore, a classifier would
                do
                > well to learn to predict that any picture with a small circle in
                the
                > middle is good, again missing the point.
                >
                > I think the reason such projects have failed in the past is
                because
                > of just such spurious correlations. The question is how to push
                > evolution to learn the deeper meaning of "good," which is so
                > profoundly slippery.
                >
                > Nevertheless, I would not say it is a lost cause. Yet it is
                > certainly a grand challenge and will not succumb easily.
                >
                > ken
                >
                >
                >
                >
                >
                >
                > ---------------------------------
                > Moody friends. Drama queens. Your life? Nope! - their life, your
                story.
                > Play Sims Stories at Yahoo! Games.
                >
              • PlaceboZA (Greg)
                In addition to emotional content, I believe it s also an attempt to explore complex issues that only a human brain (at this point in time) could have explored.
                Message 7 of 17 , Sep 17, 2007
                • 0 Attachment
                  In addition to emotional content, I believe it's also an attempt to explore complex issues that only a human brain (at this point in time) could have explored.
                  Art requires you to explore your own and others' minds to produce something extraordinary.
                  We have no known way of assimilating that amount and complexity of knowledge in such a way as to make it useful

                  Perhaps emotional content is simply a lot of excessively complex neural pathways, combined with some hormones?

                  Greg

                  On 9/17/07, Kenneth Stanley <kstanley@...> wrote:

                  The problem is that it's easy to get "crazy" just by outputting a
                  bunch of random numbers; yet that is not what artists generally seem
                  to be doing. I think what art does is show you something familiar
                  in a way you haven't seen it before. So it's kind of a variation on
                  a theme, or a new angle. It doesn't look that far outside the realm
                  of possibility that a computer could do that, although it is a very
                  tough problem.

                  I think one of the biggest difficulties with art is its emotional
                  content. While a computer may present something in a surprisingly
                  novel way, it is always somehow devoid of emotional significance,
                  since the computer is not coming from a subjective perspective the
                  way an artist would be. The shading in a scene may evoke the
                  sadness the artist remembers on a particular lonely night. Yet for
                  a computer it is merely an arbitrary filter on the colors. The time
                  when computers really infuse their works with emotional content is
                  certainly far away.

                  ken



                  --- In neat@yahoogroups.com, kelly parker <bigblockfw@...> wrote:
                  >
                  > Excuse me while I ramble.
                  >
                  > I think what you are attempting to evolve is marketability not art.
                  >
                  > True art has no mass.
                  > The physical embodiment does.
                  >
                  > Many people mistake skill for talent.
                  > This is easy enough to understand considering that cow manure can
                  be marketed as fertilizer if properly packaged.
                  > They are not the same thing.
                  >
                  > The rules for art are simple
                  > 1) There are no rules.
                  > 2) Null
                  >
                  > The rules for skill are also simple.
                  > Duplicate accepted behavior as closely as possible.
                  >
                  > The rules for marketability of "art".
                  > Herd familiarity with the basic structure.
                  > mild/mediocre/average/non offensive
                  >
                  > ref: American Idol (high skill, high marketability, little to no
                  talent....but a cute tush) The winners invariably sound like someone
                  familiar because they have the skill to reproduce the style of a
                  previously popular musician.
                  >
                  > A CD can duplicate the greatest music ever composed.
                  > A camera can duplicate the greatest paintings.
                  > A photocopier can duplicate the greatest poetry.
                  > A CNC router can duplicate the greatest sculpture.
                  > You could stretch this to say they (the machines) have skill but I
                  don't think anybody would mistake this for talent.
                  >
                  > If you study artists instead of art you'll find that a lot of them
                  have/had "mental illnesses". (that is the artists head doesn't work
                  like the herd thinks it should)
                  > Many people would argue that that "illness" is exactly what made
                  them great.
                  >
                  > So with all that mess out of the way...
                  >
                  > I can't imagine any way to code talent but an interesting
                  experiment might be to use lots of the "magic numbers" (Pi,
                  Fibonacci, etc.) in an algo. and then try to figure out how to make
                  it "crazy".
                  >
                  > Skill and marketability on the other hand should be trivial.
                  > Skill exists by default.
                  > marketability would be simply morphing graphical works that have
                  already proven popular.
                  >
                  > for what it's worth,
                  > K
                  >
                  >
                  >
                  >
                  >
                  >
                  >
                  > Kenneth Stanley <kstanley@...>

                  wrote: --- In
                  neat@yahoogroups.com, "John Arrowwood" <jarrowwx@> wrote:
                  > >
                  > > On 8/30/07, Kenneth Stanley <kstanley@> wrote:
                  > > > I agree with Peter and Jeff that it would be very difficult
                  to
                  > learn
                  > > > what is interesting, even from a relatively good data set
                  (like
                  > > > Picbreeder perhaps could provide). The problem is that
                  there
                  > are so
                  > > > many possible spurious correlations between good pictures
                  and
                  > other
                  > > > irrelevent factors that the classifier will not get the
                  deeper
                  > > > concept of what makes a good picture.
                  > > >
                  > > > Yet I agree with John in the sense that there is extremely
                  rich
                  > and
                  > > > unusual data in the Picbreeder database that may be useful
                  for
                  > > > something beyond Picbreeder itself. I just think the right
                  way
                  > to
                  > > > leverage it is tricky.
                  > > >
                  > > > The idea of complexity is also quite muddy. Unfortunately,
                  the
                  > word
                  > > > means a lot of different things. In fact, perhaps the best
                  > formal
                  > > > definition ("minimal description length") is both proven
                  > > > uncomputable in general and also does not correlate to what
                  is
                  > meant
                  > > > by the normal intuitive use of the word complexity with
                  respect
                  > to
                  > > > evolution.
                  > > >
                  > > > As evolutionary complexity goes up in the sense we usually
                  mean
                  > it,
                  > > > it is not just that the phenotype takes more bits to
                  describe.
                  > > > There is also a sense of organization, i.e. regularity, that
                  > begins
                  > > > to take hold. That is in direct opposition with the idea of
                  > minimal
                  > > > description length. In fact, what we really want is some
                  kind of
                  > > > partnership between complexity and organization.
                  Unfortunately
                  > that
                  > > > introduces yet another mysterious word: organization; it also
                  > > > introduced a partnership between two seemingly opposing
                  forces.
                  > Yet
                  > > > I think this muddiness and internal conflict is important
                  > because it
                  > > > shows that this problem is extremely slippery. It makes me
                  > wonder
                  > > > what is meant by "doing well" for humans asked to select for
                  > > > complexity.
                  > > >
                  > > > In any case, I do like the theme behind John's idea, but I
                  think
                  > > > something more sophisitcated needs to be done to capture the
                  > deeper
                  > > > reasons that certain pictures are selected.
                  > >
                  > > In classical art there are 'rules' for good works. The rule of
                  > > thirds, for example. Why do those rules exist? I would posit
                  that
                  > > those rules exist because of a mathematical relationship that
                  exists
                  > > in how the signals travel around in our brain. I suspect that a
                  > > network COULD be developed that would be able to identify the
                  pure
                  > > aesthetic merit of a particular image, because I think that at
                  that
                  > > level, aesthetics is a purely mathematical function.
                  > >
                  > > That said, I think that it would be very unlikely to be able to
                  say,
                  > > "hey, that looks like a bunny rabbit that has been
                  > borgified...cool!"
                  > > Since our choice of what to breed for is based partly on
                  aesthetics
                  > > and partly on the emotional tie to the objects that something
                  > reminds
                  > > us of, there is little chance that a controller network could
                  > totally
                  > > replace the human.
                  > >
                  > > But what if? What if a controller could be evolved which
                  learned to
                  > > recognize good image composition rules? How might such a thing
                  > extend
                  > > the usefulness of PicBreeder in general? What might it teach
                  us?
                  > > What might it say about the limits of computer controls? What
                  might
                  > > it say about US? Is there any merit at all to undertaking such
                  an
                  > > experiment? In my opinion, there is.
                  > >
                  > > -- John
                  >
                  > John, you argue persuasively for your idea, and I can see the
                  > appeal. I would not dismiss the idea that there is the
                  possibility
                  > that a classifier could evolve that detects something surprising
                  > about aesthetic value and thereby filters the search for
                  interesting
                  > images. It is an intoxicating idea and I would love to hear
                  about
                  > and see the results of any such attempt.
                  >
                  > Yet let me give you an idea about the origin of my skepticism.
                  I'm
                  > basically jaded by the underwhelming results of past such
                  efforts.
                  > For example, a friend of mine tried to evolve a neural network to
                  > detect good music in a similar approach to yours. I have heard
                  of
                  > other attempts like that one (usually they seem to be with
                  music). I
                  > realize these stories are only anecdotal, yet I draw a lesson
                  from
                  > them which shows the danger in such an approach:
                  >
                  > Evolution likes to cheat and get by easy. If you want it to
                  learn
                  > something deep and general, it tries to learn some cheap trick to
                  do
                  > almost as well without having to deal with the deep stuff. In
                  other
                  > words, like lightning, it takes the path of least resistance.
                  The
                  > problem with training to detect artistic quality is that there
                  may be
                  > predominant correlations among pictures or songs that we deem
                  good
                  > that are really not the reason they are good.
                  >
                  > For exampele, "good" songs may tend to have a lot of C notes, not
                  > because C makes a good song, but rather because it happens to
                  appear
                  > a lot in a particular genre, perhaps because it is the base note
                  of
                  > some scale. Thus, an evolved classifier will simply predict that
                  > anything with a lot of C's is good, missing the deeper reasons
                  > entirely. Or, in Picbreeder, you may notice if you look closely
                  that
                  > many (though not all) of the published images have some sort of
                  small
                  > circular formation near their center. For example this object
                  may be
                  > a nose, an eye (in a profile), a cockpit window, an iris, a
                  keyhole,
                  > a sun, a moon, a fingernail, etc. Therefore, a classifier would
                  do
                  > well to learn to predict that any picture with a small circle in
                  the
                  > middle is good, again missing the point.
                  >
                  > I think the reason such projects have failed in the past is
                  because
                  > of just such spurious correlations. The question is how to push
                  > evolution to learn the deeper meaning of "good," which is so
                  > profoundly slippery.
                  >
                  > Nevertheless, I would not say it is a lost cause. Yet it is
                  > certainly a grand challenge and will not succumb easily.
                  >
                  > ken
                  >
                  >
                  >
                  >
                  >
                  >
                  > ---------------------------------
                  > Moody friends. Drama queens. Your life? Nope! - their life, your
                  story.
                  > Play Sims Stories at Yahoo! Games.
                  >




                  --
                  `People who don't get carried away, should be`
                • Kenneth Stanley
                  ... to explore ... could have ... something ... knowledge ... neural ... Yes that very may well be what emotional content is. Yet I don t think we ll be
                  Message 8 of 17 , Sep 27, 2007
                  • 0 Attachment
                    --- In neat@yahoogroups.com, "PlaceboZA (Greg)" <placeboza@...>
                    wrote:
                    >
                    > In addition to emotional content, I believe it's also an attempt
                    to explore
                    > complex issues that only a human brain (at this point in time)
                    could have
                    > explored.
                    > Art requires you to explore your own and others' minds to produce
                    something
                    > extraordinary.
                    > We have no known way of assimilating that amount and complexity of
                    knowledge
                    > in such a way as to make it useful
                    >
                    > Perhaps emotional content is simply a lot of excessively complex
                    neural
                    > pathways, combined with some hormones?
                    >

                    Yes that very may well be what emotional content is. Yet I don't
                    think we'll be creating anything like it on computers anytime soon.
                    Still, perhaps art can be approached somehow by capturing what
                    people appreciate emotionally without actually understanding emotion
                    itself. In any case even that is extremely hard.

                    ken

                    ken


                    > Greg
                    >
                    > On 9/17/07, Kenneth Stanley <kstanley@...> wrote:
                    > >
                    > > The problem is that it's easy to get "crazy" just by
                    outputting a
                    > > bunch of random numbers; yet that is not what artists generally
                    seem
                    > > to be doing. I think what art does is show you something familiar
                    > > in a way you haven't seen it before. So it's kind of a variation
                    on
                    > > a theme, or a new angle. It doesn't look that far outside the
                    realm
                    > > of possibility that a computer could do that, although it is a
                    very
                    > > tough problem.
                    > >
                    > > I think one of the biggest difficulties with art is its emotional
                    > > content. While a computer may present something in a surprisingly
                    > > novel way, it is always somehow devoid of emotional significance,
                    > > since the computer is not coming from a subjective perspective
                    the
                    > > way an artist would be. The shading in a scene may evoke the
                    > > sadness the artist remembers on a particular lonely night. Yet
                    for
                    > > a computer it is merely an arbitrary filter on the colors. The
                    time
                    > > when computers really infuse their works with emotional content
                    is
                    > > certainly far away.
                    > >
                    > > ken
                    > >
                    > >
                    > > --- In neat@yahoogroups.com <neat%40yahoogroups.com>, kelly
                    parker
                    > > <bigblockfw@> wrote:
                    > > >
                    > > > Excuse me while I ramble.
                    > > >
                    > > > I think what you are attempting to evolve is marketability not
                    art.
                    > > >
                    > > > True art has no mass.
                    > > > The physical embodiment does.
                    > > >
                    > > > Many people mistake skill for talent.
                    > > > This is easy enough to understand considering that cow manure
                    can
                    > > be marketed as fertilizer if properly packaged.
                    > > > They are not the same thing.
                    > > >
                    > > > The rules for art are simple
                    > > > 1) There are no rules.
                    > > > 2) Null
                    > > >
                    > > > The rules for skill are also simple.
                    > > > Duplicate accepted behavior as closely as possible.
                    > > >
                    > > > The rules for marketability of "art".
                    > > > Herd familiarity with the basic structure.
                    > > > mild/mediocre/average/non offensive
                    > > >
                    > > > ref: American Idol (high skill, high marketability, little to
                    no
                    > > talent....but a cute tush) The winners invariably sound like
                    someone
                    > > familiar because they have the skill to reproduce the style of a
                    > > previously popular musician.
                    > > >
                    > > > A CD can duplicate the greatest music ever composed.
                    > > > A camera can duplicate the greatest paintings.
                    > > > A photocopier can duplicate the greatest poetry.
                    > > > A CNC router can duplicate the greatest sculpture.
                    > > > You could stretch this to say they (the machines) have skill
                    but I
                    > > don't think anybody would mistake this for talent.
                    > > >
                    > > > If you study artists instead of art you'll find that a lot of
                    them
                    > > have/had "mental illnesses". (that is the artists head doesn't
                    work
                    > > like the herd thinks it should)
                    > > > Many people would argue that that "illness" is exactly what
                    made
                    > > them great.
                    > > >
                    > > > So with all that mess out of the way...
                    > > >
                    > > > I can't imagine any way to code talent but an interesting
                    > > experiment might be to use lots of the "magic numbers" (Pi,
                    > > Fibonacci, etc.) in an algo. and then try to figure out how to
                    make
                    > > it "crazy".
                    > > >
                    > > > Skill and marketability on the other hand should be trivial.
                    > > > Skill exists by default.
                    > > > marketability would be simply morphing graphical works that
                    have
                    > > already proven popular.
                    > > >
                    > > > for what it's worth,
                    > > > K
                    > > >
                    > > >
                    > > >
                    > > >
                    > > >
                    > > >
                    > > >
                    > > > Kenneth Stanley <kstanley@>
                    > > wrote: --- In
                    > > neat@yahoogroups.com <neat%40yahoogroups.com>, "John Arrowwood"
                    > > <jarrowwx@> wrote:
                    > > > >
                    > > > > On 8/30/07, Kenneth Stanley <kstanley@> wrote:
                    > > > > > I agree with Peter and Jeff that it would be very difficult
                    > > to
                    > > > learn
                    > > > > > what is interesting, even from a relatively good data set
                    > > (like
                    > > > > > Picbreeder perhaps could provide). The problem is that
                    > > there
                    > > > are so
                    > > > > > many possible spurious correlations between good pictures
                    > > and
                    > > > other
                    > > > > > irrelevent factors that the classifier will not get the
                    > > deeper
                    > > > > > concept of what makes a good picture.
                    > > > > >
                    > > > > > Yet I agree with John in the sense that there is extremely
                    > > rich
                    > > > and
                    > > > > > unusual data in the Picbreeder database that may be useful
                    > > for
                    > > > > > something beyond Picbreeder itself. I just think the right
                    > > way
                    > > > to
                    > > > > > leverage it is tricky.
                    > > > > >
                    > > > > > The idea of complexity is also quite muddy. Unfortunately,
                    > > the
                    > > > word
                    > > > > > means a lot of different things. In fact, perhaps the best
                    > > > formal
                    > > > > > definition ("minimal description length") is both proven
                    > > > > > uncomputable in general and also does not correlate to what
                    > > is
                    > > > meant
                    > > > > > by the normal intuitive use of the word complexity with
                    > > respect
                    > > > to
                    > > > > > evolution.
                    > > > > >
                    > > > > > As evolutionary complexity goes up in the sense we usually
                    > > mean
                    > > > it,
                    > > > > > it is not just that the phenotype takes more bits to
                    > > describe.
                    > > > > > There is also a sense of organization, i.e. regularity,
                    that
                    > > > begins
                    > > > > > to take hold. That is in direct opposition with the idea of
                    > > > minimal
                    > > > > > description length. In fact, what we really want is some
                    > > kind of
                    > > > > > partnership between complexity and organization.
                    > > Unfortunately
                    > > > that
                    > > > > > introduces yet another mysterious word: organization; it
                    also
                    > > > > > introduced a partnership between two seemingly opposing
                    > > forces.
                    > > > Yet
                    > > > > > I think this muddiness and internal conflict is important
                    > > > because it
                    > > > > > shows that this problem is extremely slippery. It makes me
                    > > > wonder
                    > > > > > what is meant by "doing well" for humans asked to select
                    for
                    > > > > > complexity.
                    > > > > >
                    > > > > > In any case, I do like the theme behind John's idea, but I
                    > > think
                    > > > > > something more sophisitcated needs to be done to capture
                    the
                    > > > deeper
                    > > > > > reasons that certain pictures are selected.
                    > > > >
                    > > > > In classical art there are 'rules' for good works. The rule
                    of
                    > > > > thirds, for example. Why do those rules exist? I would posit
                    > > that
                    > > > > those rules exist because of a mathematical relationship that
                    > > exists
                    > > > > in how the signals travel around in our brain. I suspect
                    that a
                    > > > > network COULD be developed that would be able to identify the
                    > > pure
                    > > > > aesthetic merit of a particular image, because I think that
                    at
                    > > that
                    > > > > level, aesthetics is a purely mathematical function.
                    > > > >
                    > > > > That said, I think that it would be very unlikely to be able
                    to
                    > > say,
                    > > > > "hey, that looks like a bunny rabbit that has been
                    > > > borgified...cool!"
                    > > > > Since our choice of what to breed for is based partly on
                    > > aesthetics
                    > > > > and partly on the emotional tie to the objects that something
                    > > > reminds
                    > > > > us of, there is little chance that a controller network could
                    > > > totally
                    > > > > replace the human.
                    > > > >
                    > > > > But what if? What if a controller could be evolved which
                    > > learned to
                    > > > > recognize good image composition rules? How might such a
                    thing
                    > > > extend
                    > > > > the usefulness of PicBreeder in general? What might it teach
                    > > us?
                    > > > > What might it say about the limits of computer controls? What
                    > > might
                    > > > > it say about US? Is there any merit at all to undertaking
                    such
                    > > an
                    > > > > experiment? In my opinion, there is.
                    > > > >
                    > > > > -- John
                    > > >
                    > > > John, you argue persuasively for your idea, and I can see the
                    > > > appeal. I would not dismiss the idea that there is the
                    > > possibility
                    > > > that a classifier could evolve that detects something
                    surprising
                    > > > about aesthetic value and thereby filters the search for
                    > > interesting
                    > > > images. It is an intoxicating idea and I would love to hear
                    > > about
                    > > > and see the results of any such attempt.
                    > > >
                    > > > Yet let me give you an idea about the origin of my skepticism.
                    > > I'm
                    > > > basically jaded by the underwhelming results of past such
                    > > efforts.
                    > > > For example, a friend of mine tried to evolve a neural network
                    to
                    > > > detect good music in a similar approach to yours. I have heard
                    > > of
                    > > > other attempts like that one (usually they seem to be with
                    > > music). I
                    > > > realize these stories are only anecdotal, yet I draw a lesson
                    > > from
                    > > > them which shows the danger in such an approach:
                    > > >
                    > > > Evolution likes to cheat and get by easy. If you want it to
                    > > learn
                    > > > something deep and general, it tries to learn some cheap trick
                    to
                    > > do
                    > > > almost as well without having to deal with the deep stuff. In
                    > > other
                    > > > words, like lightning, it takes the path of least resistance.
                    > > The
                    > > > problem with training to detect artistic quality is that there
                    > > may be
                    > > > predominant correlations among pictures or songs that we deem
                    > > good
                    > > > that are really not the reason they are good.
                    > > >
                    > > > For exampele, "good" songs may tend to have a lot of C notes,
                    not
                    > > > because C makes a good song, but rather because it happens to
                    > > appear
                    > > > a lot in a particular genre, perhaps because it is the base
                    note
                    > > of
                    > > > some scale. Thus, an evolved classifier will simply predict
                    that
                    > > > anything with a lot of C's is good, missing the deeper reasons
                    > > > entirely. Or, in Picbreeder, you may notice if you look closely
                    > > that
                    > > > many (though not all) of the published images have some sort of
                    > > small
                    > > > circular formation near their center. For example this object
                    > > may be
                    > > > a nose, an eye (in a profile), a cockpit window, an iris, a
                    > > keyhole,
                    > > > a sun, a moon, a fingernail, etc. Therefore, a classifier would
                    > > do
                    > > > well to learn to predict that any picture with a small circle
                    in
                    > > the
                    > > > middle is good, again missing the point.
                    > > >
                    > > > I think the reason such projects have failed in the past is
                    > > because
                    > > > of just such spurious correlations. The question is how to push
                    > > > evolution to learn the deeper meaning of "good," which is so
                    > > > profoundly slippery.
                    > > >
                    > > > Nevertheless, I would not say it is a lost cause. Yet it is
                    > > > certainly a grand challenge and will not succumb easily.
                    > > >
                    > > > ken
                    > > >
                    > > >
                    > > >
                    > > >
                    > > >
                    > > >
                    > > > ---------------------------------
                    > > > Moody friends. Drama queens. Your life? Nope! - their life,
                    your
                    > > story.
                    > > > Play Sims Stories at Yahoo! Games.
                    > > >
                    > >
                    > >
                    > >
                    >
                    >
                    >
                    > --
                    > `People who don't get carried away, should be`
                    >
                  • John Arrowwood
                    ... My theory is that emotional content is a wave function of how the experience (as a wave) interferes with the expectation (memory of past experiences), also
                    Message 9 of 17 , Sep 28, 2007
                    • 0 Attachment
                      On 9/17/07, PlaceboZA (Greg) <placeboza@...> wrote:
                      > Perhaps emotional content is simply a lot of excessively complex neural pathways, combined with some hormones?

                      My theory is that emotional content is a wave function of how the
                      experience (as a wave) interferes with the expectation (memory of past
                      experiences), also as a wave. If they are consonant, they produce a
                      strong, positive emotion. If they are dissonant, it produces a strong
                      negative emotion. If my theory is correct, it would actually be
                      fairly easy to produce a computer that can experience emotions, albeit
                      ones unique to itself. Easy, that is, once you figure out how to
                      convert the sensory data into a wave function, and how to create
                      memory that works in terms of that wave function, and which produces
                      expectations in terms of that wave function. So, maybe 'easy' isn't
                      the right word... :)

                      --
                      John Arrowwood
                      John (at) Hanlons Razor (dot) com
                      John (at) Arrowwood Photography (dot) com
                      --
                      http://arrowwood.blogspot.com
                    • John Arrowwood
                      ... Image enlargement with a penalty for trying to take short-cuts. I have a printed copy of the study at home, let s see if I can find a link... There it
                      Message 10 of 17 , Sep 28, 2007
                      • 0 Attachment
                        On 9/16/07, Kenneth Stanley <kstanley@...> wrote:
                        > By the way, could you remind me what that study was that
                        > was "nothing short of stunning?" I can't remember if you cited it
                        > already, but I wonder if you have more detail on that? It
                        > definitely sounds interesting.

                        Image enlargement with a penalty for trying to take short-cuts. I
                        have a printed copy of the study at home, let's see if I can find a
                        link... There it is...

                        http://hpl.hp.com/techreports/2003/HPL-2003-26R1.pdf

                        --
                        John Arrowwood
                        John (at) Hanlons Razor (dot) com
                        John (at) Arrowwood Photography (dot) com
                        --
                        http://arrowwood.blogspot.com
                      • Kenneth Stanley
                        John, how instrumental in your view of emotion is the idea that it is a wave as opposed to some other conduit of information? What is about waves and wave
                        Message 11 of 17 , Sep 28, 2007
                        • 0 Attachment
                          John, how instrumental in your view of emotion is the idea that it is
                          a "wave" as opposed to some other conduit of information? What is
                          about waves and wave functions that makes you feel they are the right
                          level of abstraction for emotion?

                          Also, I wasn't sure, by "wave function" are you referring to quantum
                          mechanics? Or I could be misunderstanding.

                          ken

                          --- In neat@yahoogroups.com, "John Arrowwood" <jarrowwx@...> wrote:
                          >
                          > On 9/17/07, PlaceboZA (Greg) <placeboza@...> wrote:
                          > > Perhaps emotional content is simply a lot of excessively complex
                          neural pathways, combined with some hormones?
                          >
                          > My theory is that emotional content is a wave function of how the
                          > experience (as a wave) interferes with the expectation (memory of
                          past
                          > experiences), also as a wave. If they are consonant, they produce a
                          > strong, positive emotion. If they are dissonant, it produces a
                          strong
                          > negative emotion. If my theory is correct, it would actually be
                          > fairly easy to produce a computer that can experience emotions,
                          albeit
                          > ones unique to itself. Easy, that is, once you figure out how to
                          > convert the sensory data into a wave function, and how to create
                          > memory that works in terms of that wave function, and which produces
                          > expectations in terms of that wave function. So, maybe 'easy' isn't
                          > the right word... :)
                          >
                          > --
                          > John Arrowwood
                          > John (at) Hanlons Razor (dot) com
                          > John (at) Arrowwood Photography (dot) com
                          > --
                          > http://arrowwood.blogspot.com
                          >
                        • John Arrowwood
                          ... This is somewhat of a mechanistic model. The wave in question is the flowing neurochemical signal that travels through a biological neural network.
                          Message 12 of 17 , Oct 1, 2007
                          • 0 Attachment
                            On 9/28/07, Kenneth Stanley <kstanley@...> wrote:
                            > John, how instrumental in your view of emotion is the idea that it is
                            > a "wave" as opposed to some other conduit of information? What is
                            > about waves and wave functions that makes you feel they are the right
                            > level of abstraction for emotion?

                            This is somewhat of a mechanistic model. The 'wave' in question is
                            the 'flowing neurochemical signal' that travels through a biological
                            neural network.

                            Knowing that the 'wave' is implemented via neurochemicals which take
                            time to be taken back into the cells once they fire, it is easy to see
                            how these 'waves' would exhibit some of the same behaviors of our
                            classic 'waves', such as additivity. When two classic waves enter
                            into the same space, they bend their medium in such a way as to either
                            augment or cancel out. In the brain, when two signals arrive at the
                            same location in near time, the signal from the first has not yet
                            finished being taken up, so the second signal adds to what is left of
                            the first, potentially triggering that neuron to fire when the first
                            signal alone could not. Thus, they are additive.

                            Music is something we do because we like it. Every human emotion can
                            be triggered through music. Music theory tells us that we will find
                            two notes 'pleasant' if their frequencies are a fraction of two
                            relatively small integers apart from one another (I may not have
                            worded that exactly correct, but try to extract the intent. :) This
                            is best understood by graphing sine waves. The sound wave goes from
                            high density to low density in a cyclic pattern, a frequency, like a
                            sine wave. If you add another sine wave, much like mixing multiple
                            sounds together, then the one either adds to or cancels out the other
                            at any given moment. The characteristics of the resultant wave
                            determines whether or not it will be deemed 'consonant' or
                            'dissonant,' 'pleasant' or 'unpleasant.' Which it will be is well
                            understood.

                            In the brain, any regularly occurring sensation will set up a
                            'frequency following response' (you can google it) which is in essence
                            a 'standing wave' in the brain whose characteristic frequency matches
                            the frequency of the original sensation. This means that if you hear
                            one frequency in one ear (e.g. 440 hz), and another similar frequency
                            (e.g. 450 hz) in the other ear, then you will have two standing waves
                            in your brain. The one will interact with the other in a 10 hz cycle,
                            going from 'reinforcing' to 'canceling out' at a rate of 10 times a
                            second. The individual will perceive this as a 'beat frequency'
                            (google: "binaural beat" and "brainwave entrainment"). The signals
                            interact in the brain EXACTLY like they interact in your mathematical
                            model of adding sine waves together. There are significant
                            implications to this fact.

                            Add to this an interesting concept: Synesthesia. Normally, one part
                            of the brain is dedicated to processing one type of signal, and
                            different regions don't typically interact except in well defined
                            ways. However, sometimes, when the amount of real-estate that has
                            been coopted into participating in the processing of the signals
                            brings the one processing region into close proximity to another
                            region, then signals may bleed from the one to the other. This would
                            be experienced as a 'mixing up' of senses, such that you can hear a
                            color, or see a shape when you hear a sound. People who take certain
                            substances into their bloodstream experience this as a side-effect
                            (trippy!) while others naturally experience it as part of their
                            every-day experience. But everyone is capable of experiencing it if
                            they allow themselves to 'fully experience' their senses.

                            The point is that the signals from one part of the brain are not
                            incompatible with other parts of the brain. The 'language of
                            computation' of the brain is the same for the whole brain, not just
                            the audio-processing center. Thus, the rules for music are the same
                            as the rules for 'shape' or 'color' or 'scent' or 'texture.' The
                            principles that determine our emotional response to music are the same
                            rules that could be used to predict our emotional response to ANY
                            stimuli, if we can understand how the stimuli gets translated into
                            internal 'waves' of information within the brain.

                            I propose that the translation mechanism is designed or evolved so as
                            to maximize 'compatibility' between the various senses, so that one
                            can use the signals from one and mix them with the others in order to
                            find correlation and learn from them. Likewise, I propose that memory
                            is the re-creation of old wave patterns, and that our experience of
                            emotions (vs. pleasure/pain) is founded on the same exact principle,
                            just applied to the 'memory' wave as it interacts with the
                            'perception' waves (at whatever level of abstraction is appropriate).

                            As a test, take a long hard look at the things that make you happy vs.
                            the things that make you sad. You will find that the things which
                            make you unhappy are things where your belief about how they SHOULD be
                            does not line up with how you perceive things to be. But like music,
                            there is a degree of 'difference' beyond which the two don't seem to
                            interact. If they are too different, reality doesn't 'effect' you.
                            There is another region wherein they become 'dissonant' or
                            'discordant' and therefore unpleasant, and yet another 'consonant' or
                            'resonant' region which is generally experienced as joyous. The more
                            of your beliefs that are 'confirmed' by or resonant to your
                            perception, the larger the magnitude of the experience of joy. The
                            closer to resonant it is while still being off, the more dissonant it
                            is.

                            This theory completely explains why Saturday-morning cartoons are just
                            as 'believable' as live-action films, but a C.G. that is too close to
                            real but is off triggers a rejection, a negative response. It is in
                            that 'dissonant' area, where the 'waves' in the brain triggered by the
                            sensation are setting up too many expectations which the signals are
                            not meeting. It explains the difference between a good actor and a
                            bad one, because the good one more fully meets our expectations. It
                            explains our appreciation of beauty, because the mathematical
                            characteristics of the wave form that forms in our minds from our
                            sensations has some kind of positive (predictable) relationship with
                            our expectation, which is itself based on past experiences as a kind
                            of Bayesian 'average'.

                            Which is why I believe that it is possible to build a neural network
                            that could learn to recognize the characteristics of 'good' images.
                            Because ultimately, it's all math.

                            Does that make sense?

                            > Also, I wasn't sure, by "wave function" are you referring to quantum
                            > mechanics? Or I could be misunderstanding.

                            No, I'm referring to waves and how they function. Key elements are
                            the fact that a wave is technically a 'construct' or 'configuration'
                            of a medium, which 'travels' through the medium. I'm talking about
                            the 'functionality' of 'waves' and the way that mathematical
                            'functions' can describe them.

                            --
                            John Arrowwood
                            John (at) Hanlons Razor (dot) com
                            John (at) Arrowwood Photography (dot) com
                            --
                            http://arrowwood.blogspot.com
                          Your message has been successfully submitted and would be delivered to recipients shortly.