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

Re: [SeattleRobotics] Re: Language

Expand Messages
  • don clay
    I had my eyes lasered a year and a half ago. From that experience, I ve wondered how much of a role our brains play in our visual world, because long after my
    Message 1 of 18 , Oct 9, 2008
    • 0 Attachment
      I had my eyes lasered a year and a half ago. From that experience, I've wondered how
      much of a role our brains play in our visual world, because long after my eyes healed
      things were still fuzzy. Then quite quickly focus developed once again. I'm not talking
      about perception. I'm talking about seeing things clearly.

      > --- In SeattleRobotics@yahoogroups.com, "Matthew Tedder"
      > <matthewct@...> wrote:
      > >
      > > I think my model for cognition, applied to my model for neural
      > substrate might perform exceptionally as a proactive approach to
      > computer vision. This is definitely something on my list to experiment
      > with somewhere down the road. >
      >
      >
      >
      > I will have to take another look, but I hadn't noticed that your idea
      > of cognition really addressed the specific problems of perception.
      >
      > As regards the issues of so-called "movement blindness", those
      > experiments are vastly overrated. We clearly perceive [as in
      > "identifying"] mainly what we specifically are looking at [ie,
      > foveating] and attending to, and which falls into the 6-deg of so of
      > central vision, and we build up the details of any visual image over
      > successive saccades.
      >
      > What the movement blindness studies mainly illustrate is that we don't
      > capture high-resolution images into neural memory buffers for detailed
      > comparison from instant to instant, so it's not at all surprising we
      > don't notice change minutiae. In contrast, most attempts at computer
      > vision work with high-res buffers, and high-res image comparisions,
      > and that might be the chief stumbling block.
      >
      > Regards vision modeling, there are also plenty of top-down, and
      > so-called predictive approaches, but they are still not very good. I
      > recommend looking at Richard Granger's work, and not spending a lot of
      > time with Poggio/Riesenhuber or Ballard/Rao or Field/Olshausen.
      >
    • Matthew Tedder
      I think the brain is fairly good at adapting to malformations but sometimes the input just isn t of sufficient quality. I am thinking of getting that surgery,
      Message 2 of 18 , Oct 9, 2008
      • 0 Attachment
        I think the brain is fairly good at adapting to malformations but sometimes
        the input just isn't of sufficient quality. I am thinking of getting that
        surgery, too. I sure with my brain could just account for the distortions
        in light and fix the image on its own.. but.. you know.. darn.

        My guess is that it was the cornea healing that sharpened your vision over
        time.

        I read a blog article once in which the author postulated that every major
        advance in document search technology involved, not necessarily a change in
        algorithm, but an additional factor to consider. Google is the prime
        example. They derived a new factor from the existing data and
        revolutionized web search. Many subsequent Google challengers have used a
        variety of algorithms based on the same factors and they appear to give
        results not much different from Google's. Inspite of claims of
        revolutionary change, it's been nothing but moderately incremental, at best.

        So how about an algorithm to find new factors? (aka factor analysis, in
        statistics) I think neural substrate does a very good job of this.

        If incoming connections are drawn in and formed from correlate signals when
        a neuron doesn't have enough potentiation to reach threshold, and incomming
        connections are weakened and destroyed when a neuron is potentiated over
        threshold then the factors that correlate with the phenomenon represented by
        that neuron, are continuously updated for maximum relevance.

        The average number of receptors is about 1,000 but this varies greatly. The
        more, the weaker they tend to be and the less, the stronger they tend to
        be. The resting state of a cell is -70mV and threshold is -55mV. And
        throughout the nervous system, potentiation tends to stay very close to
        threshold. Note: potentiation is how much voltage is put into the soma only
        when input is received from its receptors.. and each receptor has its own
        short-term and long-term potentiation factors. All of this sums, in the
        soma.

        So, each neuron will represent the phenomenon of the correlation of other
        phenomenon (neurons firing) within "Geographic Reach"--distance verses
        signal strength (higher frequency = stronger). I once saw a real-time video
        of an axon, where little fibres were extending and rescinding and waving
        around if various directions like a kite string in turbulant wind. A neuron
        with too weak a soma will send a chemical trail for these axonial fibres to
        reach them. But an axon only extends while it is carrying an actoin
        potential (a signal). Thus, if it fire more often, simultaneously with the
        neuron drawing it in, then it'll reach there faster. Geographic distance
        is, however, always a factor. Nearer ones have the advantage in this race..
        And once the neuron is satisfied (has enough input strength), the race is
        over. So, stronger connections will come from farther away and weaker ones
        from nearer--generally (meaning, with exceptions).

        I originally presumed that the convergance these signals down a string of
        temporally connected neurons would continuously increase the signal
        strength. But my own computer simulation of this demonstrated that this was
        only sometimes true. The degree depends strongly on the geographic
        organization of the neurons and the nature of the input patterns.

        I found many interesting structures developing--and most of this had to do
        with the nature of streaming input. Signals tend to come in bursts of
        various lengths. Many structures could not develop without this. And, this
        also highlights the importance of long-term potentiation (strongly resisting
        to write more on this).

        The most influential (as I currently see it) structure is what I call the
        "Context Roll". A temporal string of neurons will repeatedly (depending on
        the length of bursts) draw connections from its earlier neurons up to its
        later neurons. This is easier to envision with more abstract (high level)
        phenomenon where LTP is stronger and the frequency of being activated is
        less, causing a more slow-motion movement along the context roll. Let's say
        you walk into your house (your house is activated) then your walk into the
        kitchen (kitchen is activated under your house) and then you take a cookie
        from the cook jar (a cookie from the kitchen of your house).

        Each element is identified both independently and also individually. Thus
        the factor of each phenomenon, as it is under each unique mix of contexts
        are identified.

        My point should be clear by now--this is an algorithm to find new factors
        from the same data.

        My biggest struggle right now, btw, is that sometimes one such context roll
        (or other structure) will inhibit another. For example, if a color red it
        cannot at the same time be blue. I my arm is following one motor pattern it
        should not simultaneously follow another. If I am paying attention to one
        thing, I cannot simultaneously pay attention to another (with caveats).

        Each receptor is either polarized (add positive voltage) or hyperpolarized
        (adds negative voltage). The question is, what makes one form one way
        verses the other? It's obvious why this is important and under what
        conditions is should--but I cannot see how. My feeling is that when this is
        understood and implemented into my code, I will see the automatic formation
        of many more structures--most importantly, the central process I spoke of in
        any earlier post.

        Matthew


        On Thu, Oct 9, 2008 at 4:58 AM, don clay <donclay@...> wrote:

        > I had my eyes lasered a year and a half ago. From that experience, I've
        > wondered how
        > much of a role our brains play in our visual world, because long after my
        > eyes healed
        > things were still fuzzy. Then quite quickly focus developed once again. I'm
        > not talking
        > about perception. I'm talking about seeing things clearly.
        >
        >
        > > --- In SeattleRobotics@yahoogroups.com<SeattleRobotics%40yahoogroups.com>,
        > "Matthew Tedder"
        > > <matthewct@...> wrote:
        > > >
        > > > I think my model for cognition, applied to my model for neural
        > > substrate might perform exceptionally as a proactive approach to
        > > computer vision. This is definitely something on my list to experiment
        > > with somewhere down the road. >
        > >
        > >
        > >
        > > I will have to take another look, but I hadn't noticed that your idea
        > > of cognition really addressed the specific problems of perception.
        > >
        > > As regards the issues of so-called "movement blindness", those
        > > experiments are vastly overrated. We clearly perceive [as in
        > > "identifying"] mainly what we specifically are looking at [ie,
        > > foveating] and attending to, and which falls into the 6-deg of so of
        > > central vision, and we build up the details of any visual image over
        > > successive saccades.
        > >
        > > What the movement blindness studies mainly illustrate is that we don't
        > > capture high-resolution images into neural memory buffers for detailed
        > > comparison from instant to instant, so it's not at all surprising we
        > > don't notice change minutiae. In contrast, most attempts at computer
        > > vision work with high-res buffers, and high-res image comparisions,
        > > and that might be the chief stumbling block.
        > >
        > > Regards vision modeling, there are also plenty of top-down, and
        > > so-called predictive approaches, but they are still not very good. I
        > > recommend looking at Richard Granger's work, and not spending a lot of
        > > time with Poggio/Riesenhuber or Ballard/Rao or Field/Olshausen.
        > >
        >
        >
        >


        [Non-text portions of this message have been removed]
      • dan michaels
        ... I ve wondered how much of a role our brains play in our visual world, because long after my eyes healed things were still fuzzy. Then quite quickly focus
        Message 3 of 18 , Oct 9, 2008
        • 0 Attachment
          --- In SeattleRobotics@yahoogroups.com, "don clay" <donclay@...> wrote:
          >
          > I had my eyes lasered a year and a half ago. From that experience,
          I've wondered how much of a role our brains play in our visual world,
          because long after my eyes healed things were still fuzzy. Then quite
          quickly focus developed once again. I'm not talking
          > about perception. I'm talking about seeing things clearly.
          >


          Brains of kids are much more malleable than those of adults, and they
          recommend kids get refractive fixes [ie, eyeglasses] very quickly, or
          the visual parts of their brains may not develop properly, and they
          can have permanent problems. This is from the classical work of
          Hubel+Wiesel from the 50s and 60s.

          My guess is, assuming you're over 6-YO [<g>], and you could see fine
          with glasses before lasing, that if you have trouble after lasing,
          then the problem is in the eyeball - cornea. [BTW, I'm not a doctor
          but sit in on "House" from time to time. LOL]
        • dan michaels
          ... You know, Matthew, I think it s a real shame you are posting so much of this stuff on this forum, and getting so little feedback. Since it is all in the
          Message 4 of 18 , Oct 9, 2008
          • 0 Attachment
            --- In SeattleRobotics@yahoogroups.com, "Matthew Tedder"
            <matthewct@...> wrote:
            >


            You know, Matthew, I think it's a real shame you are posting so much
            of this stuff on this forum, and getting so little feedback.

            Since it is all in the public domain now, how about if I post some of
            your ideas over to google comp.ai.philosophy? They could use some
            fresh infusion of ideas over there, and you'll likely get LOTS of
            feedback, one way or the other.

            BTW, I also take exception [not surprisingly] to what you've been
            saying about correlations and statistics. I am sure if that was really
            how the brain does things, then the zillions of AI/maths approaches
            along the same line would have worked better by now than they do. IOW,
            any approaches that relie mainly on search are doomed to mediocre
            performance, even though it is the classical nub of AI.


            >
            > I think the brain is fairly good at adapting to malformations but
            sometimes
            > the input just isn't of sufficient quality. I am thinking of
            getting that
            > surgery, too. I sure with my brain could just account for the
            distortions
            > in light and fix the image on its own.. but.. you know.. darn.
            >
            > My guess is that it was the cornea healing that sharpened your
            vision over
            > time.
            >
            > I read a blog article once in which the author postulated that every
            major
            > advance in document search technology involved, not necessarily a
            change in
            > algorithm, but an additional factor to consider. Google is the prime
            > example. They derived a new factor from the existing data and
            > revolutionized web search. Many subsequent Google challengers have
            used a
            > variety of algorithms based on the same factors and they appear to give
            > results not much different from Google's. Inspite of claims of
            > revolutionary change, it's been nothing but moderately incremental,
            at best.
            >
            > So how about an algorithm to find new factors? (aka factor analysis, in
            > statistics) I think neural substrate does a very good job of this.
            >
            > If incoming connections are drawn in and formed from correlate
            signals when
            > a neuron doesn't have enough potentiation to reach threshold, and
            incomming
            > connections are weakened and destroyed when a neuron is potentiated over
            > threshold then the factors that correlate with the phenomenon
            represented by
            > that neuron, are continuously updated for maximum relevance.
            >
            > The average number of receptors is about 1,000 but this varies
            greatly. The
            > more, the weaker they tend to be and the less, the stronger they tend to
            > be. The resting state of a cell is -70mV and threshold is -55mV. And
            > throughout the nervous system, potentiation tends to stay very close to
            > threshold. Note: potentiation is how much voltage is put into the
            soma only
            > when input is received from its receptors.. and each receptor has
            its own
            > short-term and long-term potentiation factors. All of this sums, in the
            > soma.
            >
            > So, each neuron will represent the phenomenon of the correlation of
            other
            > phenomenon (neurons firing) within "Geographic Reach"--distance verses
            > signal strength (higher frequency = stronger). I once saw a
            real-time video
            > of an axon, where little fibres were extending and rescinding and waving
            > around if various directions like a kite string in turbulant wind.
            A neuron
            > with too weak a soma will send a chemical trail for these axonial
            fibres to
            > reach them. But an axon only extends while it is carrying an actoin
            > potential (a signal). Thus, if it fire more often, simultaneously
            with the
            > neuron drawing it in, then it'll reach there faster. Geographic
            distance
            > is, however, always a factor. Nearer ones have the advantage in
            this race..
            > And once the neuron is satisfied (has enough input strength), the
            race is
            > over. So, stronger connections will come from farther away and
            weaker ones
            > from nearer--generally (meaning, with exceptions).
            >
            > I originally presumed that the convergance these signals down a
            string of
            > temporally connected neurons would continuously increase the signal
            > strength. But my own computer simulation of this demonstrated that
            this was
            > only sometimes true. The degree depends strongly on the geographic
            > organization of the neurons and the nature of the input patterns.
            >
            > I found many interesting structures developing--and most of this had
            to do
            > with the nature of streaming input. Signals tend to come in bursts of
            > various lengths. Many structures could not develop without this.
            And, this
            > also highlights the importance of long-term potentiation (strongly
            resisting
            > to write more on this).
            >
            > The most influential (as I currently see it) structure is what I
            call the
            > "Context Roll". A temporal string of neurons will repeatedly
            (depending on
            > the length of bursts) draw connections from its earlier neurons up
            to its
            > later neurons. This is easier to envision with more abstract (high
            level)
            > phenomenon where LTP is stronger and the frequency of being activated is
            > less, causing a more slow-motion movement along the context roll.
            Let's say
            > you walk into your house (your house is activated) then your walk
            into the
            > kitchen (kitchen is activated under your house) and then you take a
            cookie
            > from the cook jar (a cookie from the kitchen of your house).
            >
            > Each element is identified both independently and also individually.
            Thus
            > the factor of each phenomenon, as it is under each unique mix of
            contexts
            > are identified.
            >
            > My point should be clear by now--this is an algorithm to find new
            factors
            > from the same data.
            >
            > My biggest struggle right now, btw, is that sometimes one such
            context roll
            > (or other structure) will inhibit another. For example, if a color
            red it
            > cannot at the same time be blue. I my arm is following one motor
            pattern it
            > should not simultaneously follow another. If I am paying attention
            to one
            > thing, I cannot simultaneously pay attention to another (with caveats).
            >
            > Each receptor is either polarized (add positive voltage) or
            hyperpolarized
            > (adds negative voltage). The question is, what makes one form one way
            > verses the other? It's obvious why this is important and under what
            > conditions is should--but I cannot see how. My feeling is that when
            this is
            > understood and implemented into my code, I will see the automatic
            formation
            > of many more structures--most importantly, the central process I
            spoke of in
            > any earlier post.
            >
            > Matthew
            >
            >
            > On Thu, Oct 9, 2008 at 4:58 AM, don clay <donclay@...> wrote:
            >
            > > I had my eyes lasered a year and a half ago. From that
            experience, I've
            > > wondered how
            > > much of a role our brains play in our visual world, because long
            after my
            > > eyes healed
            > > things were still fuzzy. Then quite quickly focus developed once
            again. I'm
            > > not talking
            > > about perception. I'm talking about seeing things clearly.
            > >
            > >
            > > > --- In
            SeattleRobotics@yahoogroups.com<SeattleRobotics%40yahoogroups.com>,
            > > "Matthew Tedder"
            > > > <matthewct@> wrote:
            > > > >
            > > > > I think my model for cognition, applied to my model for neural
            > > > substrate might perform exceptionally as a proactive approach to
            > > > computer vision. This is definitely something on my list to
            experiment
            > > > with somewhere down the road. >
            > > >
            > > >
            > > >
            > > > I will have to take another look, but I hadn't noticed that your
            idea
            > > > of cognition really addressed the specific problems of perception.
            > > >
            > > > As regards the issues of so-called "movement blindness", those
            > > > experiments are vastly overrated. We clearly perceive [as in
            > > > "identifying"] mainly what we specifically are looking at [ie,
            > > > foveating] and attending to, and which falls into the 6-deg of so of
            > > > central vision, and we build up the details of any visual image over
            > > > successive saccades.
            > > >
            > > > What the movement blindness studies mainly illustrate is that we
            don't
            > > > capture high-resolution images into neural memory buffers for
            detailed
            > > > comparison from instant to instant, so it's not at all surprising we
            > > > don't notice change minutiae. In contrast, most attempts at computer
            > > > vision work with high-res buffers, and high-res image comparisions,
            > > > and that might be the chief stumbling block.
            > > >
            > > > Regards vision modeling, there are also plenty of top-down, and
            > > > so-called predictive approaches, but they are still not very good. I
            > > > recommend looking at Richard Granger's work, and not spending a
            lot of
            > > > time with Poggio/Riesenhuber or Ballard/Rao or Field/Olshausen.
            > > >
            > >
            > >
            > >
            >
            >
            > [Non-text portions of this message have been removed]
            >
          • Matthew Tedder
            It sounds like you are almost completely misunderstanding what I wrote. And it s feedback I reply to, each time. No shortage there. Matthew ... [Non-text
            Message 5 of 18 , Oct 9, 2008
            • 0 Attachment
              It sounds like you are almost completely misunderstanding what I wrote. And
              it's feedback I reply to, each time. No shortage there.

              Matthew


              On Thu, Oct 9, 2008 at 1:28 PM, dan michaels <oric_dan@...> wrote:

              > --- In SeattleRobotics@yahoogroups.com<SeattleRobotics%40yahoogroups.com>,
              > "Matthew Tedder"
              > <matthewct@...> wrote:
              > >
              >
              > You know, Matthew, I think it's a real shame you are posting so much
              > of this stuff on this forum, and getting so little feedback.
              >
              > Since it is all in the public domain now, how about if I post some of
              > your ideas over to google comp.ai.philosophy? They could use some
              > fresh infusion of ideas over there, and you'll likely get LOTS of
              > feedback, one way or the other.
              >
              > BTW, I also take exception [not surprisingly] to what you've been
              > saying about correlations and statistics. I am sure if that was really
              > how the brain does things, then the zillions of AI/maths approaches
              > along the same line would have worked better by now than they do. IOW,
              > any approaches that relie mainly on search are doomed to mediocre
              > performance, even though it is the classical nub of AI.
              >
              > >
              > > I think the brain is fairly good at adapting to malformations but
              > sometimes
              > > the input just isn't of sufficient quality. I am thinking of
              > getting that
              > > surgery, too. I sure with my brain could just account for the
              > distortions
              > > in light and fix the image on its own.. but.. you know.. darn.
              > >
              > > My guess is that it was the cornea healing that sharpened your
              > vision over
              > > time.
              > >
              > > I read a blog article once in which the author postulated that every
              > major
              > > advance in document search technology involved, not necessarily a
              > change in
              > > algorithm, but an additional factor to consider. Google is the prime
              > > example. They derived a new factor from the existing data and
              > > revolutionized web search. Many subsequent Google challengers have
              > used a
              > > variety of algorithms based on the same factors and they appear to give
              > > results not much different from Google's. Inspite of claims of
              > > revolutionary change, it's been nothing but moderately incremental,
              > at best.
              > >
              > > So how about an algorithm to find new factors? (aka factor analysis, in
              > > statistics) I think neural substrate does a very good job of this.
              > >
              > > If incoming connections are drawn in and formed from correlate
              > signals when
              > > a neuron doesn't have enough potentiation to reach threshold, and
              > incomming
              > > connections are weakened and destroyed when a neuron is potentiated over
              > > threshold then the factors that correlate with the phenomenon
              > represented by
              > > that neuron, are continuously updated for maximum relevance.
              > >
              > > The average number of receptors is about 1,000 but this varies
              > greatly. The
              > > more, the weaker they tend to be and the less, the stronger they tend to
              > > be. The resting state of a cell is -70mV and threshold is -55mV. And
              > > throughout the nervous system, potentiation tends to stay very close to
              > > threshold. Note: potentiation is how much voltage is put into the
              > soma only
              > > when input is received from its receptors.. and each receptor has
              > its own
              > > short-term and long-term potentiation factors. All of this sums, in the
              > > soma.
              > >
              > > So, each neuron will represent the phenomenon of the correlation of
              > other
              > > phenomenon (neurons firing) within "Geographic Reach"--distance verses
              > > signal strength (higher frequency = stronger). I once saw a
              > real-time video
              > > of an axon, where little fibres were extending and rescinding and waving
              > > around if various directions like a kite string in turbulant wind.
              > A neuron
              > > with too weak a soma will send a chemical trail for these axonial
              > fibres to
              > > reach them. But an axon only extends while it is carrying an actoin
              > > potential (a signal). Thus, if it fire more often, simultaneously
              > with the
              > > neuron drawing it in, then it'll reach there faster. Geographic
              > distance
              > > is, however, always a factor. Nearer ones have the advantage in
              > this race..
              > > And once the neuron is satisfied (has enough input strength), the
              > race is
              > > over. So, stronger connections will come from farther away and
              > weaker ones
              > > from nearer--generally (meaning, with exceptions).
              > >
              > > I originally presumed that the convergance these signals down a
              > string of
              > > temporally connected neurons would continuously increase the signal
              > > strength. But my own computer simulation of this demonstrated that
              > this was
              > > only sometimes true. The degree depends strongly on the geographic
              > > organization of the neurons and the nature of the input patterns.
              > >
              > > I found many interesting structures developing--and most of this had
              > to do
              > > with the nature of streaming input. Signals tend to come in bursts of
              > > various lengths. Many structures could not develop without this.
              > And, this
              > > also highlights the importance of long-term potentiation (strongly
              > resisting
              > > to write more on this).
              > >
              > > The most influential (as I currently see it) structure is what I
              > call the
              > > "Context Roll". A temporal string of neurons will repeatedly
              > (depending on
              > > the length of bursts) draw connections from its earlier neurons up
              > to its
              > > later neurons. This is easier to envision with more abstract (high
              > level)
              > > phenomenon where LTP is stronger and the frequency of being activated is
              > > less, causing a more slow-motion movement along the context roll.
              > Let's say
              > > you walk into your house (your house is activated) then your walk
              > into the
              > > kitchen (kitchen is activated under your house) and then you take a
              > cookie
              > > from the cook jar (a cookie from the kitchen of your house).
              > >
              > > Each element is identified both independently and also individually.
              > Thus
              > > the factor of each phenomenon, as it is under each unique mix of
              > contexts
              > > are identified.
              > >
              > > My point should be clear by now--this is an algorithm to find new
              > factors
              > > from the same data.
              > >
              > > My biggest struggle right now, btw, is that sometimes one such
              > context roll
              > > (or other structure) will inhibit another. For example, if a color
              > red it
              > > cannot at the same time be blue. I my arm is following one motor
              > pattern it
              > > should not simultaneously follow another. If I am paying attention
              > to one
              > > thing, I cannot simultaneously pay attention to another (with caveats).
              > >
              > > Each receptor is either polarized (add positive voltage) or
              > hyperpolarized
              > > (adds negative voltage). The question is, what makes one form one way
              > > verses the other? It's obvious why this is important and under what
              > > conditions is should--but I cannot see how. My feeling is that when
              > this is
              > > understood and implemented into my code, I will see the automatic
              > formation
              > > of many more structures--most importantly, the central process I
              > spoke of in
              > > any earlier post.
              > >
              > > Matthew
              > >
              > >
              > > On Thu, Oct 9, 2008 at 4:58 AM, don clay <donclay@...> wrote:
              > >
              > > > I had my eyes lasered a year and a half ago. From that
              > experience, I've
              > > > wondered how
              > > > much of a role our brains play in our visual world, because long
              > after my
              > > > eyes healed
              > > > things were still fuzzy. Then quite quickly focus developed once
              > again. I'm
              > > > not talking
              > > > about perception. I'm talking about seeing things clearly.
              > > >
              > > >
              > > > > --- In
              > SeattleRobotics@yahoogroups.com <SeattleRobotics%40yahoogroups.com>
              > <SeattleRobotics%40yahoogroups.com>,
              >
              > > > "Matthew Tedder"
              > > > > <matthewct@> wrote:
              > > > > >
              > > > > > I think my model for cognition, applied to my model for neural
              > > > > substrate might perform exceptionally as a proactive approach to
              > > > > computer vision. This is definitely something on my list to
              > experiment
              > > > > with somewhere down the road. >
              > > > >
              > > > >
              > > > >
              > > > > I will have to take another look, but I hadn't noticed that your
              > idea
              > > > > of cognition really addressed the specific problems of perception.
              > > > >
              > > > > As regards the issues of so-called "movement blindness", those
              > > > > experiments are vastly overrated. We clearly perceive [as in
              > > > > "identifying"] mainly what we specifically are looking at [ie,
              > > > > foveating] and attending to, and which falls into the 6-deg of so of
              > > > > central vision, and we build up the details of any visual image over
              > > > > successive saccades.
              > > > >
              > > > > What the movement blindness studies mainly illustrate is that we
              > don't
              > > > > capture high-resolution images into neural memory buffers for
              > detailed
              > > > > comparison from instant to instant, so it's not at all surprising we
              > > > > don't notice change minutiae. In contrast, most attempts at computer
              > > > > vision work with high-res buffers, and high-res image comparisions,
              > > > > and that might be the chief stumbling block.
              > > > >
              > > > > Regards vision modeling, there are also plenty of top-down, and
              > > > > so-called predictive approaches, but they are still not very good. I
              > > > > recommend looking at Richard Granger's work, and not spending a
              > lot of
              > > > > time with Poggio/Riesenhuber or Ballard/Rao or Field/Olshausen.
              > > > >
              > > >
              > > >
              > > >
              > >
              > >
              > > [Non-text portions of this message have been removed]
              > >
              >
              >
              >


              [Non-text portions of this message have been removed]
            • dan michaels
              ... wrote. And it s feedback I reply to, each time. No shortage there. ... You re only getting any feedback from me, and one other post. On google, there
              Message 6 of 18 , Oct 10, 2008
              • 0 Attachment
                --- In SeattleRobotics@yahoogroups.com, "Matthew Tedder"
                <matthewct@...> wrote:
                >
                > It sounds like you are almost completely misunderstanding what I
                wrote. And it's feedback I reply to, each time. No shortage there.
                >
                > Matthew
                >
                >


                You're only getting any feedback from me, and one other post. On
                google, there would probably be tons of people willing to spend the
                time to debate every issue you mention. You will notice most
                everything you write just zips by. I'm not gonna respond to all that.

                The other day you made a big deal about correlations and statistics,
                but it you can't build an intelligent system that way. No way. It
                takes a lot more. Perceptrons and connectionist nets are basically
                correlators, and they hit the wall with toy problems. Their
                limitations have been known for decades. See Minsky+Papert's
                Perceptrons book.

                All the other stuff you wrote about -70 mv and STPS and LTPs is
                standard neuro. Hebbs rule has been around since the 1940s. LTPs have
                been known since the 80s. It has something to do with the mechanism of
                learning, but no one has been able to turn that into a real "brain".
                All AI learning systems in the past have also hit the wall.

                Something is missing yet, and it's a lot more than just "learning" and
                "STPs" and "correlations", and a few general rules of cognition. I am
                sure if you look in the psychology and cog.sci literature, you will
                find dozens of schemes very similar to what you have been describing.
                I really see nothing out of the ordinary. Something is missing.

                OTOH, what you are ignoring is the sort of thing that Minsky has been
                talking about since the 1970s. Ie, hundreds of specialized processors
                that have evolved to perform certain kinds of tasks, and which are
                somehow coordinated to produce coherent output behavior. His society
                of mind concept. These things are part genetically-specified, and part
                tuned during development, and part modified by learning through life,
                and so far it's mainly still a big mystery.

                The one thing that's probably not true is that there are a few simple
                rules of cognition, like the laws of physics, that will explain
                everything, even though batallions of people have been looking for
                them for over 100 years now.


                >
                > On Thu, Oct 9, 2008 at 1:28 PM, dan michaels <oric_dan@...> wrote:
                >
                > > --- In
                SeattleRobotics@yahoogroups.com<SeattleRobotics%40yahoogroups.com>,
                > > "Matthew Tedder"
                > > <matthewct@> wrote:
                > > >
                > >
                > > You know, Matthew, I think it's a real shame you are posting so much
                > > of this stuff on this forum, and getting so little feedback.
                > >
                > > Since it is all in the public domain now, how about if I post some of
                > > your ideas over to google comp.ai.philosophy? They could use some
                > > fresh infusion of ideas over there, and you'll likely get LOTS of
                > > feedback, one way or the other.
                > >
                > > BTW, I also take exception [not surprisingly] to what you've been
                > > saying about correlations and statistics. I am sure if that was really
                > > how the brain does things, then the zillions of AI/maths approaches
                > > along the same line would have worked better by now than they do. IOW,
                > > any approaches that relie mainly on search are doomed to mediocre
                > > performance, even though it is the classical nub of AI.
                > >
                > > >
                > > > I think the brain is fairly good at adapting to malformations but
                > > sometimes
                > > > the input just isn't of sufficient quality. I am thinking of
                > > getting that
                > > > surgery, too. I sure with my brain could just account for the
                > > distortions
                > > > in light and fix the image on its own.. but.. you know.. darn.
                > > >
                > > > My guess is that it was the cornea healing that sharpened your
                > > vision over
                > > > time.
                > > >
                > > > I read a blog article once in which the author postulated that every
                > > major
                > > > advance in document search technology involved, not necessarily a
                > > change in
                > > > algorithm, but an additional factor to consider. Google is the prime
                > > > example. They derived a new factor from the existing data and
                > > > revolutionized web search. Many subsequent Google challengers have
                > > used a
                > > > variety of algorithms based on the same factors and they appear
                to give
                > > > results not much different from Google's. Inspite of claims of
                > > > revolutionary change, it's been nothing but moderately incremental,
                > > at best.
                > > >
                > > > So how about an algorithm to find new factors? (aka factor
                analysis, in
                > > > statistics) I think neural substrate does a very good job of this.
                > > >
                > > > If incoming connections are drawn in and formed from correlate
                > > signals when
                > > > a neuron doesn't have enough potentiation to reach threshold, and
                > > incomming
                > > > connections are weakened and destroyed when a neuron is
                potentiated over
                > > > threshold then the factors that correlate with the phenomenon
                > > represented by
                > > > that neuron, are continuously updated for maximum relevance.
                > > >
                > > > The average number of receptors is about 1,000 but this varies
                > > greatly. The
                > > > more, the weaker they tend to be and the less, the stronger they
                tend to
                > > > be. The resting state of a cell is -70mV and threshold is -55mV. And
                > > > throughout the nervous system, potentiation tends to stay very
                close to
                > > > threshold. Note: potentiation is how much voltage is put into the
                > > soma only
                > > > when input is received from its receptors.. and each receptor has
                > > its own
                > > > short-term and long-term potentiation factors. All of this sums,
                in the
                > > > soma.
                > > >
                > > > So, each neuron will represent the phenomenon of the correlation of
                > > other
                > > > phenomenon (neurons firing) within "Geographic Reach"--distance
                verses
                > > > signal strength (higher frequency = stronger). I once saw a
                > > real-time video
                > > > of an axon, where little fibres were extending and rescinding
                and waving
                > > > around if various directions like a kite string in turbulant wind.
                > > A neuron
                > > > with too weak a soma will send a chemical trail for these axonial
                > > fibres to
                > > > reach them. But an axon only extends while it is carrying an actoin
                > > > potential (a signal). Thus, if it fire more often, simultaneously
                > > with the
                > > > neuron drawing it in, then it'll reach there faster. Geographic
                > > distance
                > > > is, however, always a factor. Nearer ones have the advantage in
                > > this race..
                > > > And once the neuron is satisfied (has enough input strength), the
                > > race is
                > > > over. So, stronger connections will come from farther away and
                > > weaker ones
                > > > from nearer--generally (meaning, with exceptions).
                > > >
                > > > I originally presumed that the convergance these signals down a
                > > string of
                > > > temporally connected neurons would continuously increase the signal
                > > > strength. But my own computer simulation of this demonstrated that
                > > this was
                > > > only sometimes true. The degree depends strongly on the geographic
                > > > organization of the neurons and the nature of the input patterns.
                > > >
                > > > I found many interesting structures developing--and most of this had
                > > to do
                > > > with the nature of streaming input. Signals tend to come in
                bursts of
                > > > various lengths. Many structures could not develop without this.
                > > And, this
                > > > also highlights the importance of long-term potentiation (strongly
                > > resisting
                > > > to write more on this).
                > > >
                > > > The most influential (as I currently see it) structure is what I
                > > call the
                > > > "Context Roll". A temporal string of neurons will repeatedly
                > > (depending on
                > > > the length of bursts) draw connections from its earlier neurons up
                > > to its
                > > > later neurons. This is easier to envision with more abstract (high
                > > level)
                > > > phenomenon where LTP is stronger and the frequency of being
                activated is
                > > > less, causing a more slow-motion movement along the context roll.
                > > Let's say
                > > > you walk into your house (your house is activated) then your walk
                > > into the
                > > > kitchen (kitchen is activated under your house) and then you take a
                > > cookie
                > > > from the cook jar (a cookie from the kitchen of your house).
                > > >
                > > > Each element is identified both independently and also individually.
                > > Thus
                > > > the factor of each phenomenon, as it is under each unique mix of
                > > contexts
                > > > are identified.
                > > >
                > > > My point should be clear by now--this is an algorithm to find new
                > > factors
                > > > from the same data.
                > > >
                > > > My biggest struggle right now, btw, is that sometimes one such
                > > context roll
                > > > (or other structure) will inhibit another. For example, if a color
                > > red it
                > > > cannot at the same time be blue. I my arm is following one motor
                > > pattern it
                > > > should not simultaneously follow another. If I am paying attention
                > > to one
                > > > thing, I cannot simultaneously pay attention to another (with
                caveats).
                > > >
                > > > Each receptor is either polarized (add positive voltage) or
                > > hyperpolarized
                > > > (adds negative voltage). The question is, what makes one form
                one way
                > > > verses the other? It's obvious why this is important and under what
                > > > conditions is should--but I cannot see how. My feeling is that when
                > > this is
                > > > understood and implemented into my code, I will see the automatic
                > > formation
                > > > of many more structures--most importantly, the central process I
                > > spoke of in
                > > > any earlier post.
                > > >
                > > > Matthew
                > > >
                > > >
                > > > On Thu, Oct 9, 2008 at 4:58 AM, don clay <donclay@> wrote:
                > > >
                > > > > I had my eyes lasered a year and a half ago. From that
                > > experience, I've
                > > > > wondered how
                > > > > much of a role our brains play in our visual world, because long
                > > after my
                > > > > eyes healed
                > > > > things were still fuzzy. Then quite quickly focus developed once
                > > again. I'm
                > > > > not talking
                > > > > about perception. I'm talking about seeing things clearly.
                > > > >
                > > > >
                > > > > > --- In
                > > SeattleRobotics@yahoogroups.com <SeattleRobotics%40yahoogroups.com>
                > > <SeattleRobotics%40yahoogroups.com>,
                > >
                > > > > "Matthew Tedder"
                > > > > > <matthewct@> wrote:
                > > > > > >
                > > > > > > I think my model for cognition, applied to my model for neural
                > > > > > substrate might perform exceptionally as a proactive approach to
                > > > > > computer vision. This is definitely something on my list to
                > > experiment
                > > > > > with somewhere down the road. >
                > > > > >
                > > > > >
                > > > > >
                > > > > > I will have to take another look, but I hadn't noticed that your
                > > idea
                > > > > > of cognition really addressed the specific problems of
                perception.
                > > > > >
                > > > > > As regards the issues of so-called "movement blindness", those
                > > > > > experiments are vastly overrated. We clearly perceive [as in
                > > > > > "identifying"] mainly what we specifically are looking at [ie,
                > > > > > foveating] and attending to, and which falls into the 6-deg
                of so of
                > > > > > central vision, and we build up the details of any visual
                image over
                > > > > > successive saccades.
                > > > > >
                > > > > > What the movement blindness studies mainly illustrate is that we
                > > don't
                > > > > > capture high-resolution images into neural memory buffers for
                > > detailed
                > > > > > comparison from instant to instant, so it's not at all
                surprising we
                > > > > > don't notice change minutiae. In contrast, most attempts at
                computer
                > > > > > vision work with high-res buffers, and high-res image
                comparisions,
                > > > > > and that might be the chief stumbling block.
                > > > > >
                > > > > > Regards vision modeling, there are also plenty of top-down, and
                > > > > > so-called predictive approaches, but they are still not very
                good. I
                > > > > > recommend looking at Richard Granger's work, and not spending a
                > > lot of
                > > > > > time with Poggio/Riesenhuber or Ballard/Rao or Field/Olshausen.
                > > > > >
                > > > >
                > > > >
                > > > >
                > > >
                > > >
                > > > [Non-text portions of this message have been removed]
                > > >
                > >
                > >
                > >
                >
                >
                > [Non-text portions of this message have been removed]
                >
              • Matthew Tedder
                It is true I don t pay much attention to who I am responding to. I just tend to read and respond to content, not people per se. It s not debate so much that
                Message 7 of 18 , Oct 10, 2008
                • 0 Attachment
                  It is true I don't pay much attention to who I am responding to. I just
                  tend to read and respond to content, not people per se. It's not debate so
                  much that I value, in any case, but constructive criticisms--specific flaws
                  or considerations I might not have seen. I am open to change. You couldn't
                  imagine all the concepts I once felt confident in, only to learn one thing
                  that broke it completely.

                  Disclaimer: While I do have Marvin Minsky's "Society of Mind" book in my
                  bookshelf, like Stephen Pinker, I really don't give him many points for
                  making good points. I consider them both well-read but not well reasoned.

                  Importantly: A whole new world of functional structures result when streams
                  of digital signals come in (each little puff of neurotransmitter) as opposed
                  to the analog input values of perceptrons. The two don't even compare. I
                  have yet to find a pattern that cannot be captured with my model (though I
                  am always looking). I sometimes think of it as imprinting in variably
                  dimensioned clay. The spatial is converted to temporal, the temporal is
                  converted to spatial, and all correlate phenomenon that might exist should
                  be identifiable. Thus far, I haven't been able to falsify this.

                  My disagreement with you over the nature verses nurture issue is mainly
                  based on how neurons behave. The well-known mantra, "What fires together,
                  wires together." just doesn't leave much room for interconnections to have
                  been genetically determined. Also no such mechanism or event has ever been
                  found--only theorized at higher levels.

                  On the other hand, my model finds that the relative geographic locations of
                  neurons play a very substantial role in how they wire together, which is
                  clearly genetically determined--chemicals released from organizer region of
                  the endoderm cause the gene expression that determines this. So you could
                  say I have been arguing for "nature" all along, just not the way many people
                  would expect.

                  I promise you, I will publish on this. But my time is limited and it really
                  does take time.

                  Matthew

                  On Fri, Oct 10, 2008 at 1:15 PM, dan michaels <oric_dan@...> wrote:

                  > --- In SeattleRobotics@yahoogroups.com<SeattleRobotics%40yahoogroups.com>,
                  > "Matthew Tedder"
                  > <matthewct@...> wrote:
                  > >
                  > > It sounds like you are almost completely misunderstanding what I
                  > wrote. And it's feedback I reply to, each time. No shortage there.
                  > >
                  > > Matthew
                  > >
                  > >
                  >
                  > You're only getting any feedback from me, and one other post. On
                  > google, there would probably be tons of people willing to spend the
                  > time to debate every issue you mention. You will notice most
                  > everything you write just zips by. I'm not gonna respond to all that.
                  >
                  > The other day you made a big deal about correlations and statistics,
                  > but it you can't build an intelligent system that way. No way. It
                  > takes a lot more. Perceptrons and connectionist nets are basically
                  > correlators, and they hit the wall with toy problems. Their
                  > limitations have been known for decades. See Minsky+Papert's
                  > Perceptrons book.
                  >
                  > All the other stuff you wrote about -70 mv and STPS and LTPs is
                  > standard neuro. Hebbs rule has been around since the 1940s. LTPs have
                  > been known since the 80s. It has something to do with the mechanism of
                  > learning, but no one has been able to turn that into a real "brain".
                  > All AI learning systems in the past have also hit the wall.
                  >
                  > Something is missing yet, and it's a lot more than just "learning" and
                  > "STPs" and "correlations", and a few general rules of cognition. I am
                  > sure if you look in the psychology and cog.sci literature, you will
                  > find dozens of schemes very similar to what you have been describing.
                  > I really see nothing out of the ordinary. Something is missing.
                  >
                  > OTOH, what you are ignoring is the sort of thing that Minsky has been
                  > talking about since the 1970s. Ie, hundreds of specialized processors
                  > that have evolved to perform certain kinds of tasks, and which are
                  > somehow coordinated to produce coherent output behavior. His society
                  > of mind concept. These things are part genetically-specified, and part
                  > tuned during development, and part modified by learning through life,
                  > and so far it's mainly still a big mystery.
                  >
                  > The one thing that's probably not true is that there are a few simple
                  > rules of cognition, like the laws of physics, that will explain
                  > everything, even though batallions of people have been looking for
                  > them for over 100 years now.
                  >
                  > >
                  > > On Thu, Oct 9, 2008 at 1:28 PM, dan michaels <oric_dan@...> wrote:
                  > >
                  > > > --- In
                  > SeattleRobotics@yahoogroups.com <SeattleRobotics%40yahoogroups.com>
                  > <SeattleRobotics%40yahoogroups.com>,
                  > > > "Matthew Tedder"
                  > > > <matthewct@> wrote:
                  > > > >
                  > > >
                  > > > You know, Matthew, I think it's a real shame you are posting so much
                  > > > of this stuff on this forum, and getting so little feedback.
                  > > >
                  > > > Since it is all in the public domain now, how about if I post some of
                  > > > your ideas over to google comp.ai.philosophy? They could use some
                  > > > fresh infusion of ideas over there, and you'll likely get LOTS of
                  > > > feedback, one way or the other.
                  > > >
                  > > > BTW, I also take exception [not surprisingly] to what you've been
                  > > > saying about correlations and statistics. I am sure if that was really
                  > > > how the brain does things, then the zillions of AI/maths approaches
                  > > > along the same line would have worked better by now than they do. IOW,
                  > > > any approaches that relie mainly on search are doomed to mediocre
                  > > > performance, even though it is the classical nub of AI.
                  > > >
                  > > > >
                  > > > > I think the brain is fairly good at adapting to malformations but
                  > > > sometimes
                  > > > > the input just isn't of sufficient quality. I am thinking of
                  > > > getting that
                  > > > > surgery, too. I sure with my brain could just account for the
                  > > > distortions
                  > > > > in light and fix the image on its own.. but.. you know.. darn.
                  > > > >
                  > > > > My guess is that it was the cornea healing that sharpened your
                  > > > vision over
                  > > > > time.
                  > > > >
                  > > > > I read a blog article once in which the author postulated that every
                  > > > major
                  > > > > advance in document search technology involved, not necessarily a
                  > > > change in
                  > > > > algorithm, but an additional factor to consider. Google is the prime
                  > > > > example. They derived a new factor from the existing data and
                  > > > > revolutionized web search. Many subsequent Google challengers have
                  > > > used a
                  > > > > variety of algorithms based on the same factors and they appear
                  > to give
                  > > > > results not much different from Google's. Inspite of claims of
                  > > > > revolutionary change, it's been nothing but moderately incremental,
                  > > > at best.
                  > > > >
                  > > > > So how about an algorithm to find new factors? (aka factor
                  > analysis, in
                  > > > > statistics) I think neural substrate does a very good job of this.
                  > > > >
                  > > > > If incoming connections are drawn in and formed from correlate
                  > > > signals when
                  > > > > a neuron doesn't have enough potentiation to reach threshold, and
                  > > > incomming
                  > > > > connections are weakened and destroyed when a neuron is
                  > potentiated over
                  > > > > threshold then the factors that correlate with the phenomenon
                  > > > represented by
                  > > > > that neuron, are continuously updated for maximum relevance.
                  > > > >
                  > > > > The average number of receptors is about 1,000 but this varies
                  > > > greatly. The
                  > > > > more, the weaker they tend to be and the less, the stronger they
                  > tend to
                  > > > > be. The resting state of a cell is -70mV and threshold is -55mV. And
                  > > > > throughout the nervous system, potentiation tends to stay very
                  > close to
                  > > > > threshold. Note: potentiation is how much voltage is put into the
                  > > > soma only
                  > > > > when input is received from its receptors.. and each receptor has
                  > > > its own
                  > > > > short-term and long-term potentiation factors. All of this sums,
                  > in the
                  > > > > soma.
                  > > > >
                  > > > > So, each neuron will represent the phenomenon of the correlation of
                  > > > other
                  > > > > phenomenon (neurons firing) within "Geographic Reach"--distance
                  > verses
                  > > > > signal strength (higher frequency = stronger). I once saw a
                  > > > real-time video
                  > > > > of an axon, where little fibres were extending and rescinding
                  > and waving
                  > > > > around if various directions like a kite string in turbulant wind.
                  > > > A neuron
                  > > > > with too weak a soma will send a chemical trail for these axonial
                  > > > fibres to
                  > > > > reach them. But an axon only extends while it is carrying an actoin
                  > > > > potential (a signal). Thus, if it fire more often, simultaneously
                  > > > with the
                  > > > > neuron drawing it in, then it'll reach there faster. Geographic
                  > > > distance
                  > > > > is, however, always a factor. Nearer ones have the advantage in
                  > > > this race..
                  > > > > And once the neuron is satisfied (has enough input strength), the
                  > > > race is
                  > > > > over. So, stronger connections will come from farther away and
                  > > > weaker ones
                  > > > > from nearer--generally (meaning, with exceptions).
                  > > > >
                  > > > > I originally presumed that the convergance these signals down a
                  > > > string of
                  > > > > temporally connected neurons would continuously increase the signal
                  > > > > strength. But my own computer simulation of this demonstrated that
                  > > > this was
                  > > > > only sometimes true. The degree depends strongly on the geographic
                  > > > > organization of the neurons and the nature of the input patterns.
                  > > > >
                  > > > > I found many interesting structures developing--and most of this had
                  > > > to do
                  > > > > with the nature of streaming input. Signals tend to come in
                  > bursts of
                  > > > > various lengths. Many structures could not develop without this.
                  > > > And, this
                  > > > > also highlights the importance of long-term potentiation (strongly
                  > > > resisting
                  > > > > to write more on this).
                  > > > >
                  > > > > The most influential (as I currently see it) structure is what I
                  > > > call the
                  > > > > "Context Roll". A temporal string of neurons will repeatedly
                  > > > (depending on
                  > > > > the length of bursts) draw connections from its earlier neurons up
                  > > > to its
                  > > > > later neurons. This is easier to envision with more abstract (high
                  > > > level)
                  > > > > phenomenon where LTP is stronger and the frequency of being
                  > activated is
                  > > > > less, causing a more slow-motion movement along the context roll.
                  > > > Let's say
                  > > > > you walk into your house (your house is activated) then your walk
                  > > > into the
                  > > > > kitchen (kitchen is activated under your house) and then you take a
                  > > > cookie
                  > > > > from the cook jar (a cookie from the kitchen of your house).
                  > > > >
                  > > > > Each element is identified both independently and also individually.
                  > > > Thus
                  > > > > the factor of each phenomenon, as it is under each unique mix of
                  > > > contexts
                  > > > > are identified.
                  > > > >
                  > > > > My point should be clear by now--this is an algorithm to find new
                  > > > factors
                  > > > > from the same data.
                  > > > >
                  > > > > My biggest struggle right now, btw, is that sometimes one such
                  > > > context roll
                  > > > > (or other structure) will inhibit another. For example, if a color
                  > > > red it
                  > > > > cannot at the same time be blue. I my arm is following one motor
                  > > > pattern it
                  > > > > should not simultaneously follow another. If I am paying attention
                  > > > to one
                  > > > > thing, I cannot simultaneously pay attention to another (with
                  > caveats).
                  > > > >
                  > > > > Each receptor is either polarized (add positive voltage) or
                  > > > hyperpolarized
                  > > > > (adds negative voltage). The question is, what makes one form
                  > one way
                  > > > > verses the other? It's obvious why this is important and under what
                  > > > > conditions is should--but I cannot see how. My feeling is that when
                  > > > this is
                  > > > > understood and implemented into my code, I will see the automatic
                  > > > formation
                  > > > > of many more structures--most importantly, the central process I
                  > > > spoke of in
                  > > > > any earlier post.
                  > > > >
                  > > > > Matthew
                  > > > >
                  > > > >
                  > > > > On Thu, Oct 9, 2008 at 4:58 AM, don clay <donclay@> wrote:
                  > > > >
                  > > > > > I had my eyes lasered a year and a half ago. From that
                  > > > experience, I've
                  > > > > > wondered how
                  > > > > > much of a role our brains play in our visual world, because long
                  > > > after my
                  > > > > > eyes healed
                  > > > > > things were still fuzzy. Then quite quickly focus developed once
                  > > > again. I'm
                  > > > > > not talking
                  > > > > > about perception. I'm talking about seeing things clearly.
                  > > > > >
                  > > > > >
                  > > > > > > --- In
                  > > > SeattleRobotics@yahoogroups.com <SeattleRobotics%40yahoogroups.com><SeattleRobotics%
                  > 40yahoogroups.com>
                  > > > <SeattleRobotics%40yahoogroups.com>,
                  > > >
                  > > > > > "Matthew Tedder"
                  > > > > > > <matthewct@> wrote:
                  > > > > > > >
                  > > > > > > > I think my model for cognition, applied to my model for neural
                  > > > > > > substrate might perform exceptionally as a proactive approach to
                  > > > > > > computer vision. This is definitely something on my list to
                  > > > experiment
                  > > > > > > with somewhere down the road. >
                  > > > > > >
                  > > > > > >
                  > > > > > >
                  > > > > > > I will have to take another look, but I hadn't noticed that your
                  > > > idea
                  > > > > > > of cognition really addressed the specific problems of
                  > perception.
                  > > > > > >
                  > > > > > > As regards the issues of so-called "movement blindness", those
                  > > > > > > experiments are vastly overrated. We clearly perceive [as in
                  > > > > > > "identifying"] mainly what we specifically are looking at [ie,
                  > > > > > > foveating] and attending to, and which falls into the 6-deg
                  > of so of
                  > > > > > > central vision, and we build up the details of any visual
                  > image over
                  > > > > > > successive saccades.
                  > > > > > >
                  > > > > > > What the movement blindness studies mainly illustrate is that we
                  > > > don't
                  > > > > > > capture high-resolution images into neural memory buffers for
                  > > > detailed
                  > > > > > > comparison from instant to instant, so it's not at all
                  > surprising we
                  > > > > > > don't notice change minutiae. In contrast, most attempts at
                  > computer
                  > > > > > > vision work with high-res buffers, and high-res image
                  > comparisions,
                  > > > > > > and that might be the chief stumbling block.
                  > > > > > >
                  > > > > > > Regards vision modeling, there are also plenty of top-down, and
                  > > > > > > so-called predictive approaches, but they are still not very
                  > good. I
                  > > > > > > recommend looking at Richard Granger's work, and not spending a
                  > > > lot of
                  > > > > > > time with Poggio/Riesenhuber or Ballard/Rao or Field/Olshausen.
                  > > > > > >
                  > > > > >
                  > > > > >
                  > > > > >
                  > > > >
                  > > > >
                  > > > > [Non-text portions of this message have been removed]
                  > > > >
                  > > >
                  > > >
                  > > >
                  > >
                  > >
                  > > [Non-text portions of this message have been removed]
                  > >
                  >
                  >
                  >


                  [Non-text portions of this message have been removed]
                • dan michaels
                  ... in my bookshelf, like Stephen Pinker, I really don t give him many points for making good points. I consider them both well-read but not well reasoned.
                  Message 8 of 18 , Oct 10, 2008
                  • 0 Attachment
                    --- In SeattleRobotics@yahoogroups.com, "Matthew Tedder"
                    <matthewct@...> wrote:
                    >

                    >
                    > Disclaimer: While I do have Marvin Minsky's "Society of Mind" book
                    in my bookshelf, like Stephen Pinker, I really don't give him many
                    points for making good points. I consider them both well-read but not
                    well reasoned.
                    >


                    There is not much denying that the human cortex probably has more than
                    100 separate processing areas, with many cross-connections. Even the
                    cat has about 65 cortical areas, and over 1100 white matter pathways
                    connecting them together [cf, Gerald Edelman's books].

                    Minsky tries to make some sense of this with his ideas, but they are
                    only very lossely based on the anatomy. He's still enamored of the
                    idea of finding the "few, simple laws of cognition", despite also
                    pushing the many processor idea. I find his stuff a little too schizoid.

                    Pinker [whose ideas I rather dislike] is an extremist evolutionary
                    psychologist, apparently way out on the far [nature] end of the
                    nature-nurture spectrum. From my readng, ANY extremist position is
                    likely to be wrong.


                    >
                    > Importantly: A whole new world of functional structures result when
                    streams of digital signals come in (each little puff of
                    neurotransmitter) as opposed to the analog input values of
                    perceptrons. The two don't even compare.
                    >


                    Again, everything you say can be disagreed with ad infinitum. In fact,
                    the dendritic tree of each neuron is a huge complex ANALOG processor
                    that integrates in complex ways the [what you call digital] synaptic
                    inputs.



                    >
                    >I have yet to find a pattern that cannot be captured with my model
                    (though I am always looking). I sometimes think of it as imprinting
                    in variably dimensioned clay. The spatial is converted to temporal,
                    the temporal is converted to spatial, and all correlate phenomenon
                    that might exist should be identifiable. Thus far, I haven't been
                    able to falsify this.
                    >


                    It's undoubtedly true that brain activity consists of spatial-temporal
                    patterns. This has been known since the advent of the EEG machine
                    almost 100 years ago. But with 100B neurons arranged in 100 or more
                    cortical areas, and each neuron receiving 1000s of inputs from others,
                    no one has been able to get a handle on the complexity.


                    >
                    > My disagreement with you over the nature verses nurture issue is
                    mainly based on how neurons behave. The well-known mantra, "What
                    fires together,
                    > wires together." just doesn't leave much room for interconnections
                    to have been genetically determined. Also no such mechanism or event
                    has ever been found--only theorized at higher levels.
                    >


                    I've already told you several times where this is wrong. First read
                    Marcus' book.

                    http://www.google.com/custom?q=gary+marcus+birth+mind

                    Then realize, it's never either/or, it's always both/and. Consider
                    what is determined by genetics, what is tuned during development, and
                    what is modified by learning. It takes all three.

                    I'm done here. No point in just repeating the same thing over and
                    over. Gakk.



                    >
                    > On the other hand, my model finds that the relative geographic
                    locations of neurons play a very substantial role in how they wire
                    together, which is clearly genetically determined--chemicals released
                    from organizer region of the endoderm cause the gene expression that
                    determines this. So you could say I have been arguing for "nature"
                    all along, just not the way many people would expect.
                    >


                    Ideas like this have been beat to death in the literature since at
                    least the 1970s.

                    Also, if you really want to understand how gene expression factors in
                    development, look at both Marcus' book above, and also Sean Carroll's
                    book ...

                    http://www.google.com/custom?q=sean+carroll+endless+forms



                    >
                    > I promise you, I will publish on this. But my time is limited and
                    it really
                    > does take time.
                    >
                    > Matthew
                    >
                    > On Fri, Oct 10, 2008 at 1:15 PM, dan michaels <oric_dan@...> wrote:
                    >
                    > > --- In
                    SeattleRobotics@yahoogroups.com<SeattleRobotics%40yahoogroups.com>,
                    > > "Matthew Tedder"
                    > > <matthewct@> wrote:
                    > > >
                    > > > It sounds like you are almost completely misunderstanding what I
                    > > wrote. And it's feedback I reply to, each time. No shortage there.
                    > > >
                    > > > Matthew
                    > > >
                    > > >
                    > >
                    > > You're only getting any feedback from me, and one other post. On
                    > > google, there would probably be tons of people willing to spend the
                    > > time to debate every issue you mention. You will notice most
                    > > everything you write just zips by. I'm not gonna respond to all that.
                    > >
                    > > The other day you made a big deal about correlations and statistics,
                    > > but it you can't build an intelligent system that way. No way. It
                    > > takes a lot more. Perceptrons and connectionist nets are basically
                    > > correlators, and they hit the wall with toy problems. Their
                    > > limitations have been known for decades. See Minsky+Papert's
                    > > Perceptrons book.
                    > >
                    > > All the other stuff you wrote about -70 mv and STPS and LTPs is
                    > > standard neuro. Hebbs rule has been around since the 1940s. LTPs have
                    > > been known since the 80s. It has something to do with the mechanism of
                    > > learning, but no one has been able to turn that into a real "brain".
                    > > All AI learning systems in the past have also hit the wall.
                    > >
                    > > Something is missing yet, and it's a lot more than just "learning" and
                    > > "STPs" and "correlations", and a few general rules of cognition. I am
                    > > sure if you look in the psychology and cog.sci literature, you will
                    > > find dozens of schemes very similar to what you have been describing.
                    > > I really see nothing out of the ordinary. Something is missing.
                    > >
                    > > OTOH, what you are ignoring is the sort of thing that Minsky has been
                    > > talking about since the 1970s. Ie, hundreds of specialized processors
                    > > that have evolved to perform certain kinds of tasks, and which are
                    > > somehow coordinated to produce coherent output behavior. His society
                    > > of mind concept. These things are part genetically-specified, and part
                    > > tuned during development, and part modified by learning through life,
                    > > and so far it's mainly still a big mystery.
                    > >
                    > > The one thing that's probably not true is that there are a few simple
                    > > rules of cognition, like the laws of physics, that will explain
                    > > everything, even though batallions of people have been looking for
                    > > them for over 100 years now.
                    > >
                    > > >
                    > > > On Thu, Oct 9, 2008 at 1:28 PM, dan michaels <oric_dan@> wrote:
                    > > >
                    > > > > --- In
                    > > SeattleRobotics@yahoogroups.com <SeattleRobotics%40yahoogroups.com>
                    > > <SeattleRobotics%40yahoogroups.com>,
                    > > > > "Matthew Tedder"
                    > > > > <matthewct@> wrote:
                    > > > > >
                    > > > >
                    > > > > You know, Matthew, I think it's a real shame you are posting
                    so much
                    > > > > of this stuff on this forum, and getting so little feedback.
                    > > > >
                    > > > > Since it is all in the public domain now, how about if I post
                    some of
                    > > > > your ideas over to google comp.ai.philosophy? They could use some
                    > > > > fresh infusion of ideas over there, and you'll likely get LOTS of
                    > > > > feedback, one way or the other.
                    > > > >
                    > > > > BTW, I also take exception [not surprisingly] to what you've been
                    > > > > saying about correlations and statistics. I am sure if that
                    was really
                    > > > > how the brain does things, then the zillions of AI/maths
                    approaches
                    > > > > along the same line would have worked better by now than they
                    do. IOW,
                    > > > > any approaches that relie mainly on search are doomed to mediocre
                    > > > > performance, even though it is the classical nub of AI.
                    > > > >
                    > > > > >
                    > > > > > I think the brain is fairly good at adapting to
                    malformations but
                    > > > > sometimes
                    > > > > > the input just isn't of sufficient quality. I am thinking of
                    > > > > getting that
                    > > > > > surgery, too. I sure with my brain could just account for the
                    > > > > distortions
                    > > > > > in light and fix the image on its own.. but.. you know.. darn.
                    > > > > >
                    > > > > > My guess is that it was the cornea healing that sharpened your
                    > > > > vision over
                    > > > > > time.
                    > > > > >
                    > > > > > I read a blog article once in which the author postulated
                    that every
                    > > > > major
                    > > > > > advance in document search technology involved, not
                    necessarily a
                    > > > > change in
                    > > > > > algorithm, but an additional factor to consider. Google is
                    the prime
                    > > > > > example. They derived a new factor from the existing data and
                    > > > > > revolutionized web search. Many subsequent Google
                    challengers have
                    > > > > used a
                    > > > > > variety of algorithms based on the same factors and they appear
                    > > to give
                    > > > > > results not much different from Google's. Inspite of claims of
                    > > > > > revolutionary change, it's been nothing but moderately
                    incremental,
                    > > > > at best.
                    > > > > >
                    > > > > > So how about an algorithm to find new factors? (aka factor
                    > > analysis, in
                    > > > > > statistics) I think neural substrate does a very good job of
                    this.
                    > > > > >
                    > > > > > If incoming connections are drawn in and formed from correlate
                    > > > > signals when
                    > > > > > a neuron doesn't have enough potentiation to reach
                    threshold, and
                    > > > > incomming
                    > > > > > connections are weakened and destroyed when a neuron is
                    > > potentiated over
                    > > > > > threshold then the factors that correlate with the phenomenon
                    > > > > represented by
                    > > > > > that neuron, are continuously updated for maximum relevance.
                    > > > > >
                    > > > > > The average number of receptors is about 1,000 but this varies
                    > > > > greatly. The
                    > > > > > more, the weaker they tend to be and the less, the stronger they
                    > > tend to
                    > > > > > be. The resting state of a cell is -70mV and threshold is
                    -55mV. And
                    > > > > > throughout the nervous system, potentiation tends to stay very
                    > > close to
                    > > > > > threshold. Note: potentiation is how much voltage is put
                    into the
                    > > > > soma only
                    > > > > > when input is received from its receptors.. and each
                    receptor has
                    > > > > its own
                    > > > > > short-term and long-term potentiation factors. All of this sums,
                    > > in the
                    > > > > > soma.
                    > > > > >
                    > > > > > So, each neuron will represent the phenomenon of the
                    correlation of
                    > > > > other
                    > > > > > phenomenon (neurons firing) within "Geographic Reach"--distance
                    > > verses
                    > > > > > signal strength (higher frequency = stronger). I once saw a
                    > > > > real-time video
                    > > > > > of an axon, where little fibres were extending and rescinding
                    > > and waving
                    > > > > > around if various directions like a kite string in turbulant
                    wind.
                    > > > > A neuron
                    > > > > > with too weak a soma will send a chemical trail for these
                    axonial
                    > > > > fibres to
                    > > > > > reach them. But an axon only extends while it is carrying an
                    actoin
                    > > > > > potential (a signal). Thus, if it fire more often,
                    simultaneously
                    > > > > with the
                    > > > > > neuron drawing it in, then it'll reach there faster. Geographic
                    > > > > distance
                    > > > > > is, however, always a factor. Nearer ones have the advantage in
                    > > > > this race..
                    > > > > > And once the neuron is satisfied (has enough input
                    strength), the
                    > > > > race is
                    > > > > > over. So, stronger connections will come from farther away and
                    > > > > weaker ones
                    > > > > > from nearer--generally (meaning, with exceptions).
                    > > > > >
                    > > > > > I originally presumed that the convergance these signals down a
                    > > > > string of
                    > > > > > temporally connected neurons would continuously increase the
                    signal
                    > > > > > strength. But my own computer simulation of this
                    demonstrated that
                    > > > > this was
                    > > > > > only sometimes true. The degree depends strongly on the
                    geographic
                    > > > > > organization of the neurons and the nature of the input
                    patterns.
                    > > > > >
                    > > > > > I found many interesting structures developing--and most of
                    this had
                    > > > > to do
                    > > > > > with the nature of streaming input. Signals tend to come in
                    > > bursts of
                    > > > > > various lengths. Many structures could not develop without this.
                    > > > > And, this
                    > > > > > also highlights the importance of long-term potentiation
                    (strongly
                    > > > > resisting
                    > > > > > to write more on this).
                    > > > > >
                    > > > > > The most influential (as I currently see it) structure is what I
                    > > > > call the
                    > > > > > "Context Roll". A temporal string of neurons will repeatedly
                    > > > > (depending on
                    > > > > > the length of bursts) draw connections from its earlier
                    neurons up
                    > > > > to its
                    > > > > > later neurons. This is easier to envision with more abstract
                    (high
                    > > > > level)
                    > > > > > phenomenon where LTP is stronger and the frequency of being
                    > > activated is
                    > > > > > less, causing a more slow-motion movement along the context
                    roll.
                    > > > > Let's say
                    > > > > > you walk into your house (your house is activated) then your
                    walk
                    > > > > into the
                    > > > > > kitchen (kitchen is activated under your house) and then you
                    take a
                    > > > > cookie
                    > > > > > from the cook jar (a cookie from the kitchen of your house).
                    > > > > >
                    > > > > > Each element is identified both independently and also
                    individually.
                    > > > > Thus
                    > > > > > the factor of each phenomenon, as it is under each unique mix of
                    > > > > contexts
                    > > > > > are identified.
                    > > > > >
                    > > > > > My point should be clear by now--this is an algorithm to
                    find new
                    > > > > factors
                    > > > > > from the same data.
                    > > > > >
                    > > > > > My biggest struggle right now, btw, is that sometimes one such
                    > > > > context roll
                    > > > > > (or other structure) will inhibit another. For example, if a
                    color
                    > > > > red it
                    > > > > > cannot at the same time be blue. I my arm is following one motor
                    > > > > pattern it
                    > > > > > should not simultaneously follow another. If I am paying
                    attention
                    > > > > to one
                    > > > > > thing, I cannot simultaneously pay attention to another (with
                    > > caveats).
                    > > > > >
                    > > > > > Each receptor is either polarized (add positive voltage) or
                    > > > > hyperpolarized
                    > > > > > (adds negative voltage). The question is, what makes one form
                    > > one way
                    > > > > > verses the other? It's obvious why this is important and
                    under what
                    > > > > > conditions is should--but I cannot see how. My feeling is
                    that when
                    > > > > this is
                    > > > > > understood and implemented into my code, I will see the
                    automatic
                    > > > > formation
                    > > > > > of many more structures--most importantly, the central process I
                    > > > > spoke of in
                    > > > > > any earlier post.
                    > > > > >
                    > > > > > Matthew
                    > > > > >
                    > > > > >
                    > > > > > On Thu, Oct 9, 2008 at 4:58 AM, don clay <donclay@> wrote:
                    > > > > >
                    > > > > > > I had my eyes lasered a year and a half ago. From that
                    > > > > experience, I've
                    > > > > > > wondered how
                    > > > > > > much of a role our brains play in our visual world,
                    because long
                    > > > > after my
                    > > > > > > eyes healed
                    > > > > > > things were still fuzzy. Then quite quickly focus
                    developed once
                    > > > > again. I'm
                    > > > > > > not talking
                    > > > > > > about perception. I'm talking about seeing things clearly.
                    > > > > > >
                    > > > > > >
                    > > > > > > > --- In
                    > > > > SeattleRobotics@yahoogroups.com
                    <SeattleRobotics%40yahoogroups.com><SeattleRobotics%
                    > > 40yahoogroups.com>
                    > > > > <SeattleRobotics%40yahoogroups.com>,
                    > > > >
                    > > > > > > "Matthew Tedder"
                    > > > > > > > <matthewct@> wrote:
                    > > > > > > > >
                    > > > > > > > > I think my model for cognition, applied to my model
                    for neural
                    > > > > > > > substrate might perform exceptionally as a proactive
                    approach to
                    > > > > > > > computer vision. This is definitely something on my list to
                    > > > > experiment
                    > > > > > > > with somewhere down the road. >
                    > > > > > > >
                    > > > > > > >
                    > > > > > > >
                    > > > > > > > I will have to take another look, but I hadn't noticed
                    that your
                    > > > > idea
                    > > > > > > > of cognition really addressed the specific problems of
                    > > perception.
                    > > > > > > >
                    > > > > > > > As regards the issues of so-called "movement blindness",
                    those
                    > > > > > > > experiments are vastly overrated. We clearly perceive [as in
                    > > > > > > > "identifying"] mainly what we specifically are looking
                    at [ie,
                    > > > > > > > foveating] and attending to, and which falls into the 6-deg
                    > > of so of
                    > > > > > > > central vision, and we build up the details of any visual
                    > > image over
                    > > > > > > > successive saccades.
                    > > > > > > >
                    > > > > > > > What the movement blindness studies mainly illustrate is
                    that we
                    > > > > don't
                    > > > > > > > capture high-resolution images into neural memory
                    buffers for
                    > > > > detailed
                    > > > > > > > comparison from instant to instant, so it's not at all
                    > > surprising we
                    > > > > > > > don't notice change minutiae. In contrast, most attempts at
                    > > computer
                    > > > > > > > vision work with high-res buffers, and high-res image
                    > > comparisions,
                    > > > > > > > and that might be the chief stumbling block.
                    > > > > > > >
                    > > > > > > > Regards vision modeling, there are also plenty of
                    top-down, and
                    > > > > > > > so-called predictive approaches, but they are still not very
                    > > good. I
                    > > > > > > > recommend looking at Richard Granger's work, and not
                    spending a
                    > > > > lot of
                    > > > > > > > time with Poggio/Riesenhuber or Ballard/Rao or
                    Field/Olshausen.
                    > > > > > > >
                    > > > > > >
                    > > > > > >
                    > > > > > >
                    > > > > >
                    > > > > >
                    > > > > > [Non-text portions of this message have been removed]
                    > > > > >
                    > > > >
                    > > > >
                    > > > >
                    > > >
                    > > >
                    > > > [Non-text portions of this message have been removed]
                    > > >
                    > >
                    > >
                    > >
                    >
                    >
                    > [Non-text portions of this message have been removed]
                    >
                  • Matthew Tedder
                    I understand that these topics have been beat to death over a long stretch of time. I ve experienced it, too, over a long period of time. I am generally
                    Message 9 of 18 , Oct 10, 2008
                    • 0 Attachment
                      I understand that these topics have been beat to death over a long stretch
                      of time. I've experienced it, too, over a long period of time. I am
                      generally inclined to think it takes all three, as you suggest--but I keep
                      finding it more the way I am suggesting. I don't doubt what you say, in
                      general. But I need specifics without which I cannot change the picture I
                      have. Also, what I have has been really successful for me, thus far.

                      I will take a look at the books you mention.

                      Matthew

                      On Fri, Oct 10, 2008 at 2:45 PM, dan michaels <oric_dan@...> wrote:

                      > --- In SeattleRobotics@yahoogroups.com<SeattleRobotics%40yahoogroups.com>,
                      > "Matthew Tedder"
                      > <matthewct@...> wrote:
                      > >
                      >
                      > >
                      > > Disclaimer: While I do have Marvin Minsky's "Society of Mind" book
                      > in my bookshelf, like Stephen Pinker, I really don't give him many
                      > points for making good points. I consider them both well-read but not
                      > well reasoned.
                      > >
                      >
                      > There is not much denying that the human cortex probably has more than
                      > 100 separate processing areas, with many cross-connections. Even the
                      > cat has about 65 cortical areas, and over 1100 white matter pathways
                      > connecting them together [cf, Gerald Edelman's books].
                      >
                      > Minsky tries to make some sense of this with his ideas, but they are
                      > only very lossely based on the anatomy. He's still enamored of the
                      > idea of finding the "few, simple laws of cognition", despite also
                      > pushing the many processor idea. I find his stuff a little too schizoid.
                      >
                      > Pinker [whose ideas I rather dislike] is an extremist evolutionary
                      > psychologist, apparently way out on the far [nature] end of the
                      > nature-nurture spectrum. From my readng, ANY extremist position is
                      > likely to be wrong.
                      >
                      > >
                      > > Importantly: A whole new world of functional structures result when
                      > streams of digital signals come in (each little puff of
                      > neurotransmitter) as opposed to the analog input values of
                      > perceptrons. The two don't even compare.
                      > >
                      >
                      > Again, everything you say can be disagreed with ad infinitum. In fact,
                      > the dendritic tree of each neuron is a huge complex ANALOG processor
                      > that integrates in complex ways the [what you call digital] synaptic
                      > inputs.
                      >
                      > >
                      > >I have yet to find a pattern that cannot be captured with my model
                      > (though I am always looking). I sometimes think of it as imprinting
                      > in variably dimensioned clay. The spatial is converted to temporal,
                      > the temporal is converted to spatial, and all correlate phenomenon
                      > that might exist should be identifiable. Thus far, I haven't been
                      > able to falsify this.
                      > >
                      >
                      > It's undoubtedly true that brain activity consists of spatial-temporal
                      > patterns. This has been known since the advent of the EEG machine
                      > almost 100 years ago. But with 100B neurons arranged in 100 or more
                      > cortical areas, and each neuron receiving 1000s of inputs from others,
                      > no one has been able to get a handle on the complexity.
                      >
                      > >
                      > > My disagreement with you over the nature verses nurture issue is
                      > mainly based on how neurons behave. The well-known mantra, "What
                      > fires together,
                      > > wires together." just doesn't leave much room for interconnections
                      > to have been genetically determined. Also no such mechanism or event
                      > has ever been found--only theorized at higher levels.
                      > >
                      >
                      > I've already told you several times where this is wrong. First read
                      > Marcus' book.
                      >
                      > http://www.google.com/custom?q=gary+marcus+birth+mind
                      >
                      > Then realize, it's never either/or, it's always both/and. Consider
                      > what is determined by genetics, what is tuned during development, and
                      > what is modified by learning. It takes all three.
                      >
                      > I'm done here. No point in just repeating the same thing over and
                      > over. Gakk.
                      >
                      > >
                      > > On the other hand, my model finds that the relative geographic
                      > locations of neurons play a very substantial role in how they wire
                      > together, which is clearly genetically determined--chemicals released
                      > from organizer region of the endoderm cause the gene expression that
                      > determines this. So you could say I have been arguing for "nature"
                      > all along, just not the way many people would expect.
                      > >
                      >
                      > Ideas like this have been beat to death in the literature since at
                      > least the 1970s.
                      >
                      > Also, if you really want to understand how gene expression factors in
                      > development, look at both Marcus' book above, and also Sean Carroll's
                      > book ...
                      >
                      > http://www.google.com/custom?q=sean+carroll+endless+forms
                      >
                      > >
                      > > I promise you, I will publish on this. But my time is limited and
                      > it really
                      > > does take time.
                      > >
                      > > Matthew
                      > >
                      > > On Fri, Oct 10, 2008 at 1:15 PM, dan michaels <oric_dan@...> wrote:
                      > >
                      > > > --- In
                      > SeattleRobotics@yahoogroups.com <SeattleRobotics%40yahoogroups.com>
                      > <SeattleRobotics%40yahoogroups.com>,
                      > > > "Matthew Tedder"
                      > > > <matthewct@> wrote:
                      > > > >
                      > > > > It sounds like you are almost completely misunderstanding what I
                      > > > wrote. And it's feedback I reply to, each time. No shortage there.
                      > > > >
                      > > > > Matthew
                      > > > >
                      > > > >
                      > > >
                      > > > You're only getting any feedback from me, and one other post. On
                      > > > google, there would probably be tons of people willing to spend the
                      > > > time to debate every issue you mention. You will notice most
                      > > > everything you write just zips by. I'm not gonna respond to all that.
                      > > >
                      > > > The other day you made a big deal about correlations and statistics,
                      > > > but it you can't build an intelligent system that way. No way. It
                      > > > takes a lot more. Perceptrons and connectionist nets are basically
                      > > > correlators, and they hit the wall with toy problems. Their
                      > > > limitations have been known for decades. See Minsky+Papert's
                      > > > Perceptrons book.
                      > > >
                      > > > All the other stuff you wrote about -70 mv and STPS and LTPs is
                      > > > standard neuro. Hebbs rule has been around since the 1940s. LTPs have
                      > > > been known since the 80s. It has something to do with the mechanism of
                      > > > learning, but no one has been able to turn that into a real "brain".
                      > > > All AI learning systems in the past have also hit the wall.
                      > > >
                      > > > Something is missing yet, and it's a lot more than just "learning" and
                      > > > "STPs" and "correlations", and a few general rules of cognition. I am
                      > > > sure if you look in the psychology and cog.sci literature, you will
                      > > > find dozens of schemes very similar to what you have been describing.
                      > > > I really see nothing out of the ordinary. Something is missing.
                      > > >
                      > > > OTOH, what you are ignoring is the sort of thing that Minsky has been
                      > > > talking about since the 1970s. Ie, hundreds of specialized processors
                      > > > that have evolved to perform certain kinds of tasks, and which are
                      > > > somehow coordinated to produce coherent output behavior. His society
                      > > > of mind concept. These things are part genetically-specified, and part
                      > > > tuned during development, and part modified by learning through life,
                      > > > and so far it's mainly still a big mystery.
                      > > >
                      > > > The one thing that's probably not true is that there are a few simple
                      > > > rules of cognition, like the laws of physics, that will explain
                      > > > everything, even though batallions of people have been looking for
                      > > > them for over 100 years now.
                      > > >
                      > > > >
                      > > > > On Thu, Oct 9, 2008 at 1:28 PM, dan michaels <oric_dan@> wrote:
                      > > > >
                      > > > > > --- In
                      > > > SeattleRobotics@yahoogroups.com <SeattleRobotics%40yahoogroups.com><SeattleRobotics%
                      > 40yahoogroups.com>
                      > > > <SeattleRobotics%40yahoogroups.com>,
                      > > > > > "Matthew Tedder"
                      > > > > > <matthewct@> wrote:
                      > > > > > >
                      > > > > >
                      > > > > > You know, Matthew, I think it's a real shame you are posting
                      > so much
                      > > > > > of this stuff on this forum, and getting so little feedback.
                      > > > > >
                      > > > > > Since it is all in the public domain now, how about if I post
                      > some of
                      > > > > > your ideas over to google comp.ai.philosophy? They could use some
                      > > > > > fresh infusion of ideas over there, and you'll likely get LOTS of
                      > > > > > feedback, one way or the other.
                      > > > > >
                      > > > > > BTW, I also take exception [not surprisingly] to what you've been
                      > > > > > saying about correlations and statistics. I am sure if that
                      > was really
                      > > > > > how the brain does things, then the zillions of AI/maths
                      > approaches
                      > > > > > along the same line would have worked better by now than they
                      > do. IOW,
                      > > > > > any approaches that relie mainly on search are doomed to mediocre
                      > > > > > performance, even though it is the classical nub of AI.
                      > > > > >
                      > > > > > >
                      > > > > > > I think the brain is fairly good at adapting to
                      > malformations but
                      > > > > > sometimes
                      > > > > > > the input just isn't of sufficient quality. I am thinking of
                      > > > > > getting that
                      > > > > > > surgery, too. I sure with my brain could just account for the
                      > > > > > distortions
                      > > > > > > in light and fix the image on its own.. but.. you know.. darn.
                      > > > > > >
                      > > > > > > My guess is that it was the cornea healing that sharpened your
                      > > > > > vision over
                      > > > > > > time.
                      > > > > > >
                      > > > > > > I read a blog article once in which the author postulated
                      > that every
                      > > > > > major
                      > > > > > > advance in document search technology involved, not
                      > necessarily a
                      > > > > > change in
                      > > > > > > algorithm, but an additional factor to consider. Google is
                      > the prime
                      > > > > > > example. They derived a new factor from the existing data and
                      > > > > > > revolutionized web search. Many subsequent Google
                      > challengers have
                      > > > > > used a
                      > > > > > > variety of algorithms based on the same factors and they appear
                      > > > to give
                      > > > > > > results not much different from Google's. Inspite of claims of
                      > > > > > > revolutionary change, it's been nothing but moderately
                      > incremental,
                      > > > > > at best.
                      > > > > > >
                      > > > > > > So how about an algorithm to find new factors? (aka factor
                      > > > analysis, in
                      > > > > > > statistics) I think neural substrate does a very good job of
                      > this.
                      > > > > > >
                      > > > > > > If incoming connections are drawn in and formed from correlate
                      > > > > > signals when
                      > > > > > > a neuron doesn't have enough potentiation to reach
                      > threshold, and
                      > > > > > incomming
                      > > > > > > connections are weakened and destroyed when a neuron is
                      > > > potentiated over
                      > > > > > > threshold then the factors that correlate with the phenomenon
                      > > > > > represented by
                      > > > > > > that neuron, are continuously updated for maximum relevance.
                      > > > > > >
                      > > > > > > The average number of receptors is about 1,000 but this varies
                      > > > > > greatly. The
                      > > > > > > more, the weaker they tend to be and the less, the stronger they
                      > > > tend to
                      > > > > > > be. The resting state of a cell is -70mV and threshold is
                      > -55mV. And
                      > > > > > > throughout the nervous system, potentiation tends to stay very
                      > > > close to
                      > > > > > > threshold. Note: potentiation is how much voltage is put
                      > into the
                      > > > > > soma only
                      > > > > > > when input is received from its receptors.. and each
                      > receptor has
                      > > > > > its own
                      > > > > > > short-term and long-term potentiation factors. All of this sums,
                      > > > in the
                      > > > > > > soma.
                      > > > > > >
                      > > > > > > So, each neuron will represent the phenomenon of the
                      > correlation of
                      > > > > > other
                      > > > > > > phenomenon (neurons firing) within "Geographic Reach"--distance
                      > > > verses
                      > > > > > > signal strength (higher frequency = stronger). I once saw a
                      > > > > > real-time video
                      > > > > > > of an axon, where little fibres were extending and rescinding
                      > > > and waving
                      > > > > > > around if various directions like a kite string in turbulant
                      > wind.
                      > > > > > A neuron
                      > > > > > > with too weak a soma will send a chemical trail for these
                      > axonial
                      > > > > > fibres to
                      > > > > > > reach them. But an axon only extends while it is carrying an
                      > actoin
                      > > > > > > potential (a signal). Thus, if it fire more often,
                      > simultaneously
                      > > > > > with the
                      > > > > > > neuron drawing it in, then it'll reach there faster. Geographic
                      > > > > > distance
                      > > > > > > is, however, always a factor. Nearer ones have the advantage in
                      > > > > > this race..
                      > > > > > > And once the neuron is satisfied (has enough input
                      > strength), the
                      > > > > > race is
                      > > > > > > over. So, stronger connections will come from farther away and
                      > > > > > weaker ones
                      > > > > > > from nearer--generally (meaning, with exceptions).
                      > > > > > >
                      > > > > > > I originally presumed that the convergance these signals down a
                      > > > > > string of
                      > > > > > > temporally connected neurons would continuously increase the
                      > signal
                      > > > > > > strength. But my own computer simulation of this
                      > demonstrated that
                      > > > > > this was
                      > > > > > > only sometimes true. The degree depends strongly on the
                      > geographic
                      > > > > > > organization of the neurons and the nature of the input
                      > patterns.
                      > > > > > >
                      > > > > > > I found many interesting structures developing--and most of
                      > this had
                      > > > > > to do
                      > > > > > > with the nature of streaming input. Signals tend to come in
                      > > > bursts of
                      > > > > > > various lengths. Many structures could not develop without this.
                      > > > > > And, this
                      > > > > > > also highlights the importance of long-term potentiation
                      > (strongly
                      > > > > > resisting
                      > > > > > > to write more on this).
                      > > > > > >
                      > > > > > > The most influential (as I currently see it) structure is what I
                      > > > > > call the
                      > > > > > > "Context Roll". A temporal string of neurons will repeatedly
                      > > > > > (depending on
                      > > > > > > the length of bursts) draw connections from its earlier
                      > neurons up
                      > > > > > to its
                      > > > > > > later neurons. This is easier to envision with more abstract
                      > (high
                      > > > > > level)
                      > > > > > > phenomenon where LTP is stronger and the frequency of being
                      > > > activated is
                      > > > > > > less, causing a more slow-motion movement along the context
                      > roll.
                      > > > > > Let's say
                      > > > > > > you walk into your house (your house is activated) then your
                      > walk
                      > > > > > into the
                      > > > > > > kitchen (kitchen is activated under your house) and then you
                      > take a
                      > > > > > cookie
                      > > > > > > from the cook jar (a cookie from the kitchen of your house).
                      > > > > > >
                      > > > > > > Each element is identified both independently and also
                      > individually.
                      > > > > > Thus
                      > > > > > > the factor of each phenomenon, as it is under each unique mix of
                      > > > > > contexts
                      > > > > > > are identified.
                      > > > > > >
                      > > > > > > My point should be clear by now--this is an algorithm to
                      > find new
                      > > > > > factors
                      > > > > > > from the same data.
                      > > > > > >
                      > > > > > > My biggest struggle right now, btw, is that sometimes one such
                      > > > > > context roll
                      > > > > > > (or other structure) will inhibit another. For example, if a
                      > color
                      > > > > > red it
                      > > > > > > cannot at the same time be blue. I my arm is following one motor
                      > > > > > pattern it
                      > > > > > > should not simultaneously follow another. If I am paying
                      > attention
                      > > > > > to one
                      > > > > > > thing, I cannot simultaneously pay attention to another (with
                      > > > caveats).
                      > > > > > >
                      > > > > > > Each receptor is either polarized (add positive voltage) or
                      > > > > > hyperpolarized
                      > > > > > > (adds negative voltage). The question is, what makes one form
                      > > > one way
                      > > > > > > verses the other? It's obvious why this is important and
                      > under what
                      > > > > > > conditions is should--but I cannot see how. My feeling is
                      > that when
                      > > > > > this is
                      > > > > > > understood and implemented into my code, I will see the
                      > automatic
                      > > > > > formation
                      > > > > > > of many more structures--most importantly, the central process I
                      > > > > > spoke of in
                      > > > > > > any earlier post.
                      > > > > > >
                      > > > > > > Matthew
                      > > > > > >
                      > > > > > >
                      > > > > > > On Thu, Oct 9, 2008 at 4:58 AM, don clay <donclay@> wrote:
                      > > > > > >
                      > > > > > > > I had my eyes lasered a year and a half ago. From that
                      > > > > > experience, I've
                      > > > > > > > wondered how
                      > > > > > > > much of a role our brains play in our visual world,
                      > because long
                      > > > > > after my
                      > > > > > > > eyes healed
                      > > > > > > > things were still fuzzy. Then quite quickly focus
                      > developed once
                      > > > > > again. I'm
                      > > > > > > > not talking
                      > > > > > > > about perception. I'm talking about seeing things clearly.
                      > > > > > > >
                      > > > > > > >
                      > > > > > > > > --- In
                      > > > > > SeattleRobotics@yahoogroups.com<SeattleRobotics%40yahoogroups.com>
                      > <SeattleRobotics%40yahoogroups.com><SeattleRobotics%
                      > > > 40yahoogroups.com>
                      > > > > > <SeattleRobotics%40yahoogroups.com>,
                      > > > > >
                      > > > > > > > "Matthew Tedder"
                      > > > > > > > > <matthewct@> wrote:
                      > > > > > > > > >
                      > > > > > > > > > I think my model for cognition, applied to my model
                      > for neural
                      > > > > > > > > substrate might perform exceptionally as a proactive
                      > approach to
                      > > > > > > > > computer vision. This is definitely something on my list to
                      > > > > > experiment
                      > > > > > > > > with somewhere down the road. >
                      > > > > > > > >
                      > > > > > > > >
                      > > > > > > > >
                      > > > > > > > > I will have to take another look, but I hadn't noticed
                      > that your
                      > > > > > idea
                      > > > > > > > > of cognition really addressed the specific problems of
                      > > > perception.
                      > > > > > > > >
                      > > > > > > > > As regards the issues of so-called "movement blindness",
                      > those
                      > > > > > > > > experiments are vastly overrated. We clearly perceive [as in
                      > > > > > > > > "identifying"] mainly what we specifically are looking
                      > at [ie,
                      > > > > > > > > foveating] and attending to, and which falls into the 6-deg
                      > > > of so of
                      > > > > > > > > central vision, and we build up the details of any visual
                      > > > image over
                      > > > > > > > > successive saccades.
                      > > > > > > > >
                      > > > > > > > > What the movement blindness studies mainly illustrate is
                      > that we
                      > > > > > don't
                      > > > > > > > > capture high-resolution images into neural memory
                      > buffers for
                      > > > > > detailed
                      > > > > > > > > comparison from instant to instant, so it's not at all
                      > > > surprising we
                      > > > > > > > > don't notice change minutiae. In contrast, most attempts at
                      > > > computer
                      > > > > > > > > vision work with high-res buffers, and high-res image
                      > > > comparisions,
                      > > > > > > > > and that might be the chief stumbling block.
                      > > > > > > > >
                      > > > > > > > > Regards vision modeling, there are also plenty of
                      > top-down, and
                      > > > > > > > > so-called predictive approaches, but they are still not very
                      > > > good. I
                      > > > > > > > > recommend looking at Richard Granger's work, and not
                      > spending a
                      > > > > > lot of
                      > > > > > > > > time with Poggio/Riesenhuber or Ballard/Rao or
                      > Field/Olshausen.
                      > > > > > > > >
                      > > > > > > >
                      > > > > > > >
                      > > > > > > >
                      > > > > > >
                      > > > > > >
                      > > > > > > [Non-text portions of this message have been removed]
                      > > > > > >
                      > > > > >
                      > > > > >
                      > > > > >
                      > > > >
                      > > > >
                      > > > > [Non-text portions of this message have been removed]
                      > > > >
                      > > >
                      > > >
                      > > >
                      > >
                      > >
                      > > [Non-text portions of this message have been removed]
                      > >
                      >
                      >
                      >


                      [Non-text portions of this message have been removed]
                    Your message has been successfully submitted and would be delivered to recipients shortly.