Critique #1 Re: A Critique of Prof. Hubert Dreyfus' "Why Heideggerian AI failed"- a call for comments
Isaac
the
particular words at issue by enclosing them between "***" characters. I'll
also include citations in the paper when helpful. I seek (intelligent and
informed) technical/theoretical critique or feedback from anyone on this
particular issue. Ask/email me for a copy of the paper if you are
interested in
the context and details.
2nd critique, on his page 12, line 4:
"Heidegger's important insight is not that, when we solve problems, we
sometimes make use of representational equipment outside our bodies, but
that being-in-the-world is more basic than thinking and solving
problems;that it is not representational at all. That is, when we are
coping at our best, ***we are drawn in by solicitations and respond directly
to them, so that the distinction between us and our equipment--between inner
and outer-vanishes***#1 As Heidegger sums it up:
I live in the understanding of writing, illuminating, going-in-and-out, and
the like. More precisely: as Dasein I am -- in speaking, going, and
understanding -- an act of understanding dealing-with. My being in the
world is nothing other than this already-operating-with-understanding in
this mode of being.[ii]
Heidegger and Merleau-Ponty's understanding of embedded embodied coping,
then, is not that the mind is sometimes extended into the world but rather
that all such problem solving is derivative, that in our most basic way of
being, that is, as absorbed skillful copers, we are not minds at all but one
with the world. Heidegger sticks to the phenomenon, when he makes the
strange-sounding claim that, in its most basic way of being, "Dasein is its
world existingly."[iii]
When you stop thinking that mind is what characterizes us most basically
but, rather, that most basically we are absorbed copers, the inner/outer
distinction becomes problematic. There's no easily askable question as to
whether the absorbed coping is in me or in the world. According to
Heidegger, intentional content isn't in the mind, nor in some 3rd realm (as
it is for Husserl), nor in the world; it isn't anywhere. It's an embodied
way of being-towards. Thus for a Heideggerian, all forms of cognitivist
externalism presuppose a more basic existential externalism where even to
speak of "externalism" is misleading since such talk presupposes a contrast
with the internal. Compared to this genuinely Heideggerian view,
***extended-mind externalism is contrived, trivial, and irrelevant***#2.
--------------------------------------------------------------------------------
[i] As Heidegger puts it: "The self must forget itself if, lost in the world
of equipment, it is to be able 'actually' to go to work and manipulate
something." Being and Time, 405.
[ii] Logic, 146. It's important to realize that when he uses the term
"understanding," Heidegger explains (with a little help from the translator)
that he means a kind of know-how:
In German we say that someone can vorstehen something-literally, stand in
front of or ahead of it, that is, stand at its head, administer, manage,
preside over it. This is equivalent to saying that he versteht sich darauf,
understands in the sense of being skilled or expert at it, has the know-how
of it. (Martin Heidegger, The Basic Problems of Phenomenology, A.
Hofstadter, Trans. Bloomington: Indian University Press, 1982, 276.)
[iii] Being and Time, 416. To make sense of this slogan, it's important to
be clear that Heidegger distinguishes the human world from the physical
universe.
--------------------------------------------------------------------------------
My critique #1:
seems that the "distinction between us and our equipment... is vanished" is
just describing the unconscious automation process that takes over body
functions and relieves the conscious mind to be unaware that its equipment
was drawn into responding to solicitations. This in many ways seems to just
be alluding to the domain of our unconscious being that responds like
dominos that fall automatically in response to many contextual
solicitations. I do not see how this all makes a solid argument that
conscious thought is unified and inseparable from "our equipment" (i.e.,
body). At best this is a very weak, if not completely flawed, logic in
inferring that our sense (act) of being in the world "is not
representational at all". The text that appears to clarify this assertion
just seems to be a string of conclusory declarations without a solid logical
foundation. Even a plausible syllogism would be helpful here."
My critique #2:
is not the Heideggerian view requiring this unity between the mind and the
world result in a "contrived, trivial, and irrelevant" world representation
scheme in people when the events in the world are so far beyond a person's
ability to cope (relative to there internal representation/value system)
that they just end up contriving a trivial and irrelevant internal world
that is just projected onto a "best fit/nearest neighbor" of a
representation that they can cope with. In this way, there is no absorbed
coping because it requires a perfect and accurate absorption scheme between
our mind (inner) and the world (outer) that does not exist and cannot be
magically created, even biologically. If you ignore this aspect of the
Heideggerian view then what you end up with is nothing much more than an
"ignorance is bliss" cognitive model that is not too different from what you
say is wrong with Brook's approach. That is, your portrayal of the
Heideggerian view of absorbed coping would exactly model the thinking and
representation behavior of insects, which certainly is not the conscious,
cognitive model of humans. Thus, this Heideggerian view of absorbed coping
is either insufficient to describe the human condition or it renders
indistinguishable insects from humans; either way it does not seem to
uniquely capture the behavior at the level of human consciousness and is,
thus, flawed at best. That is, if this Heideggerian view of absorbed
coping equally applies to any animals or insects then it is not really
helpful to modeling or shedding light on higher human intellectual
behavior, which, of course, is the sole subject/goal of AI. Moreover, this
"perfect absorption" is a complete illusion and in practice will only exist
in the most predictable and simple situations. From another angle, how is
this Heideggerian view of absorbed coping much different from the standard
psychological model of projection where our internal model/representation is
simply projected onto the world (or a subset frame of it) and we just trick
ourselves into believing that we are completely and accurately absorbed with
the true essence of the frame problem. this Heideggerian view of absorbed
coping seems to much more fit the unconscious aspects of the human
condition, which is more insect/animal like. This all seems to be logically
flawed and/or a very weak foundation for grandiose conclusions about what
philosophical approach/model is needed to solve the frame problem and human
consciousness. Maybe I am missing something critical here that can make
sense of it. Please clarify the logic.
Any thoughts on this issue?
Ariel B.
"Isaac" <gro...@sonic.net> wrote in message
news:491d60f6$0$33588$742e...@news.sonic.net...
> All,
>
> I have critiqued in great detail a recent write paper by Prof. Hubert
> Dreyfus entitled "Why Heideggerian AI Failed and how Fixing it would
> Require
> making it more Heideggerian" . I can email a copy of it to whom ever is
> interested. For his bio, see:
> http://socrates.berkeley.edu/~hdreyfus/
>
> I want to stimulate discussion on this topic by posting my critiques
> little
> by little and getting comments from the AI community on the news groups.
> However, before I start I want to get a feel for how many know of his work
> and/or would be interested in an intellectual debate for and against his
> many anti-AI positions.
>
> I hope many will respond to this posting with interest so I can begin
> posting each part of this paper I find issues with and my reasoned
> critique
> for others to comment on.
>
> Thanks,
> Ariel-
>
>
Isaac
will
resonate will everyone so pick and choose what you find interesting to
debate pro/con
and I will defend any of my comments.
I will post the paragraph(s) I have a comment about, and highlight the
particular words at issue by enclosing them between "***" characters. I'll
also include citations in the paper when helpful. I seek (intelligent and
informed) technical/theoretical critique or feedback from anyone on this
particular issue.
See page 11, line 20 of his paper where it says:
"I agree that it is time for a positive account of Heideggerian AI and of an
underlying Heideggerian neuroscience, but I think Wheeler is the one looking
in the wrong place. Merely by supposing that Heidegger is concerned with
problem solving and action oriented representations, Wheeler's project
reflects not a step beyond Agre but a regression to aspects of pre-Brooks
GOFAI. Heidegger, indeed, claims that that skillful coping is basic, but he
is also clear that, all coping takes place on the background coping he calls
being-in-the-world that doesn't involve any form of representation at all.
see: Michael Wheeler, Reconstructing the Cognitive World, 222-223.
Wheeler's cognitivist misreading of Heidegger leads him to overestimate the
importance of Andy Clark's and David Chalmers' attempt to free us from the
Cartesian idea that the mind is essentially inner by pointing out that in
thinking we sometimes make use of external artifacts like pencil, paper, and
computers.[i] Unfortunately, this argument for the extended mind preserves
the Cartesian assumption that our basic way of relating to the world is by
using propositional representations such as beliefs and memories whether
they are in the mind or in notebooks in the world. In effect, while Brooks
happily dispenses with representations where coping is concerned, all
Chalmers, Clark, and Wheeler give us as a supposedly radical new
Heideggerian approach to the human way of being in the world is to note that
memories and beliefs are not necessarily inner entities and that,
***therefore, thinking bridges the distinction between inner and outer
representations.*** "
My Critique:
"Assuming that by "thinking" you mean conscious thought, I cannot see how
thinking is a bridge that necessarily follows from memories/beliefs not
being solely inner entities. It seems to me that inner and outer
representations can be bridged without thought. Isn't this what occurs in
an unconscious (reflex) reaction to a complex external even, which is an
automatic bridge and generates a thoughtful, usually accurate response but
often before we even have a chance to think about it. Inner/outer
representations seems semantically vague here. Also, cannot conscious
thought can endeavor itself with in purely inner or out representations
without ever bridging them? I guess, it is the "therefore" that gives me
pause here."
Any thoughts on this issue?
Ariel B.
"Alpha" <omegaz...@yahoo.com> wrote in message
news:f8636636-6ee8-4c92...@k24g2000pri.googlegroups.com...
On Nov 14, 9:46 pm, "Isaac" <gro...@sonic.net> wrote:
> All,
>>
>> I have critiqued in great detail a recent write paper by Prof. Hubert
>> Dreyfus entitled "Why Heideggerian AI Failed and how Fixing it would
>> Require
>> making it more Heideggerian" . I can email a copy of it to whom ever is
>> interested.
>
>Please send a copy to omegaz...@yahoo.com Ariel. Thanks!
>
>Why don't you post a summary of that paper and your key critique
>points here.
Isaac
will resonate will everyone so pick and choose what you find interesting to
debate pro/con and I will defend any of my comments.
I will post the paragraph(s) I have a comment about, and highlight the
particular words at issue by enclosing them between "***" characters. I'll
also include citations in the paper when helpful. I seek (intelligent and
informed) technical/theoretical critique or feedback from anyone on the
issue(s) presented/raised.
See page 12, line 28:
VI. What Motivates Embedded/embodied Coping?
But why is Dasein called to cope at all? According to Heidegger, we are
constantly solicited to improve our familiarity with the world. Five years
before the publication of Being and Time he wrote:
Caring takes the form of a looking around and seeing, and as this
circumspective caring it is at the same time . concerned about developing
its circumspection, that is, about securing and expanding its familiarity
with the objects of its dealings. [i]
This pragmatic perspective is developed by Merleau-Ponty, and by Samuel
Todes.[ii] These heirs to Heidegger's account of familiarity and coping
describe how an organism, animal or human, interacts with what is
objectively speaking the meaningless physical universe in such a way as to
cope with an environment organized in terms of that organism's need to find
its way around. All such coping beings are motivated to get a more and more
refined and secure sense of the specific objects of their dealings.
According to Merleau-Ponty:
My body is geared into the world when my perception presents me with a
spectacle as varied and as clearly articulated as possible...[iii]
In short, in our skilled activity we are drawn to move so as to achieve a
better and better grip on our situation. For this movement towards maximal
grip to take place one doesn't need a mental representation of one's goal
nor any problem solving, as would a GOFAI robot. ***Rather, acting is
experienced as a steady flow of skillful activity in response to the
situation. When one's situation deviates from some optimal body-environment
gestalt, one's activity takes one closer to that optimum and thereby
relieves the "tension" of the deviation. ***[asb1] ***One does not need to
know what the optimum is in order to move towards it. One's body is simply
drawn to lower the tension.*** [asb2]
***That is, if things are going well and I am gaining an optimal grip on the
world, I simple respond to the solicitation to move towards an even better
grip and, if things are going badly, I experience a pull back towards the
norm.*** [asb3] If it seems that much of the time we don't experience any
such pull, Merleau-Ponty would no doubt respond that the sensitivity to
deviation is nonetheless guiding one's coping, just as an airport radio
beacon doesn't give a warning signal unless the plane strays off course, and
then, let us suppose, the plane gets a signal whose intensity corresponds to
how far off course it is and the intensity of the signal diminishes as it
approaches getting back on course. The silence that accompanies being on
course doesn't mean the beacon isn't continually guiding the plane.
Likewise, the absence of felt tension in perception doesn't mean we aren't
***being directed by a solicitation***[asb4] .
As Merleau-Ponty puts it: ***"Our body is not an object for an 'I think', it
is a grouping of lived-through meanings that moves towards its
equilibrium***[asb5] ."[iv] Equilibrium being Merleau-Ponty's name for the
zero gradient of steady successful coping. Moreover, normally, we do not
arrive at equilibrium and stop there but are immediately taken over by a new
solicitation.
--------------------------------------------------------------------------------
MY CRITIQUES indexed by my initials "ASB" followed by the number of my
comment above:
[asb1]this sounds a lot like dominoes combined with cognitive dissonance
theory at a more sensory level. This "skillful coping" is much like a
cascade of dominoes that automatically fall in right pattern in response to
a certain stimulus. It seems that you try to avoid the need for a mental
representation by making the stimulus automatically trigger a response that
is perfectly adapted instead of going through an abstracted mental
representation of the event and using that model to calculate the best
response. Such notions of the right gestalt to every sensory
event/situation is not much different than a well trained neural network
that takes sensory inputs that are spread to all neural nodes, and depending
upon the landscape of trained weights, the learned pattern of output
response it "automatically" achieved. It would seem that such a behavioral
system has no mental representation; however, I would disagree. For
example, any sufficiently complex decision landscape will have to deal with
an uncertain range, combination, and timing of sensory inputs, so this
logically necessitates that all decision making elements will necessarily
have to represent various abstracted aspects of the sensory pattern that was
learned to be most key to achieve the desired output and error level. This
representation of the various abstracted aspects is effectively the gestalt
that each (or multiple) neuron has learned alone or in combination with
others. Hence, because a finite system that must deal with high uncertainty
must contain abstracted representations then all such systems can be thought
of as having mental representations. It is just when the task at hand is
more low level the represented abstractions seem like automatic,
unspeakable, gestalts as opposed to highly abstracted models that would tend
to occur for processing events that are more disconnected from the sensory
moment. In this way, I don't see how you can refute mental representations.
I contend, for example, that your mental representations can be thought of
as sensory abstractions. So, are you saying that your absorbed coping
occurs is without abstractions? I assert that the process of abstraction
necessarily builds a (parametric) model of the observation (i.e., a mental
representation, even if distributed) and if your system does not abstract it
cannot cope with uncertainty or variation of sensory patterns. Gestalts are
nothing more that highly effective abstractions (even heuristics) that treat
a whole class of sensory situations with an appropriate response- this is an
abstracted rule at its best. Thus, I do not see how your absorbed coping
paradyne avoids creating mental representations as I define above. It would
be helpful if you would explain your logic in terms of a practical (or even
plausible) computational system. Your Freeman-based chaotic neural network
model did not seem to address/resolve the issues I question above.
[asb2]now you have me completely confused. It is impossible to generate an
error signal (i.e., "lower the tension") without comparing against some
model of the expected or desired event/result. The fact you define it this
way seem to me to require a mental representation that represents an
abstract (representative or ideal) configuration by which the sensory inputs
are measured against (a la Plato's parallel universe of ideal
representations for all objects) to generate an error signal (tension) in a
control loop. Even the most simple negative feedback control loop requires
an abstract model of the problem domain being acted upon. It is just that
this domain model is encoded into the control loop wiring, time constants,
and gain design, which are in effect an embedded mental representation that
was created by a human and encoded into hardware or software. The body
(organic systems) must do the same thing, they just have automatic
algorithms to create the models that design the control loop. This does not
avoid mental representations it just distributes them system wide instead of
concentrating them in a centralized control scheme that most think of as a
mental representation process.
[asb3]I contend that in any practical system, there will always have to be
a model by which sensory inputs need to be compared to thereby generating an
error signal for the control loop. This model is a mental representation
even if it is hardwired. Genetic algorithms all require a fitness function.
Whether that fitness function (i.e., " optimal body-environment gestalt ")
is created by a high order mental process or by natural selection is
immaterial. The fitness function is a model, and it an Ingram of a mental
representation. I am eager to hear your arguments that practically and
logically motivate otherwise.
[asb4]it seems to me that " being directed by a solicitation " does not
contradict a guiding mental representation being present as well. My prior
comments, in my mind, avoid these two being mutually exclusive as you seem
to contend. It seems you are trying to say that a river (i.e., conscious
action) must flow down a mountain and the way it finds its minimal energy
path is dictated by natural forces (i.e., "the mind/body gestalt"), thus you
conclude that no centralized mental control/representation system is guiding
the river to its "intended" destination at the mountain base. Consciousness
is all about observing the rivers path down the mountain and adjusting the
(mental) landscape to guide the river (conscious action) to a desired
result/location. Gravity may "solicit" water down the mountain, but that
does not necessarily negate there being other forces/systems acting on the
river and shaping the landscape/path the water is to take. It seems to me
that your solicitation is like gravity, and you are concluding that there
are no mental representations because absorbed coping is guiding the river
directed by such solicitation (gravity). This is just one concrete example
(of many) of how I see your coping/solicitation/no-representations scheme as
logically and fundamentally flawed.
[asb5]It is curious how you are in effect defining "I think" as a
centralized, disconnected system from the body and environment. How is your
" lived-through meanings that moves towards its equilibrium [asb5]."
different from a classical neural network which learns meaning (as
distributed abstractions) and uses them to "move towards its equilibrium"?
There is certainly no "I think" in a classical neural network, and it
behaves as you philosophically describe. Again, if your proposed cognitive
model does not distinguish from systems that we know cannot be conscious
then it must be refined to exclude such insufficient systems as being within
the set of definitions (behaviors) required of a sufficient human cognitive
system. Otherwise, such proposed behavioral/architectural models are not
really useful to shed light on how the human mind works.
Citations made in white paper section above:
--------------------------------------------------------------------------------
[i] Martin Heidegger, Phenomenological Interpretations in Connection with
Aristotle, in Supplements: From the Earliest Essays to Being and Time and
Beyond, John Van Buren, Ed., State University of New York Press, 2002, 115.
My italics.
This away of putting the source of significance covers both animals and
people. By the time he published Being and Time, however, Heidegger was
interested exclusively in the special kind of significance found in the
world opened up by human beings who are defined by the stand they take on
their own being. We might call this meaning. In this paper I'm putting the
question of uniquely human meaning aside to concentrate on the sort of
significance we share with animals.
[ii] See, Samuel Todes, Body and World, Cambridge, MA: The MIT Press, 2001.
Todes goes beyond Merleau-Ponty in showing how our world-disclosing
perceptual experience is structured by the structure of our bodies.
Merleau-Ponty never tells us what our bodies are actually like and how their
structure affects our experience. Todes points out that our body has a
front/back and up/down orientation. It moves forward more easily than
backward, and can successfully cope only with what is in front of it. He
then describes how, in order to explore our surrounding world and orient
ourselves in it, we have to balance ourselves within a vertical field that
we do not produce, be effectively directed in a circumstantial field (facing
one aspect of that field rather than another), and appropriately set to
respond to the specific thing we are encountering within that field. For
Todes, then, perceptual receptivity is an embodied, normative, skilled
accomplishment, in response to our need to orient ourselves in the world.
Clearly, this is a kind of holistic background coping is not done for a
reason.
[iii] Merleau-Ponty, Phenomenology of Perception, 250. (Trans. Modified.)
[iv] Ibid, 153. (My italics.)
Immortalist
> Reminder: I will post the paragraph(s) I have a comment about, and highlight
> the
> particular words at issue by enclosing them between "***" characters. I'll
> also include citations in the paper when helpful. I seek (intelligent and
> informed) technical/theoretical critique or feedback from anyone on this
> particular issue. Ask/email me for a copy of the paper if you are
> interested in
> the context and details.
>
> 2nd critique, on his page 12, line 4:
> "Heidegger's important insight is not that, when we solve problems, we
> sometimes make use of representational equipment outside our bodies, but
> that being-in-the-world is more basic than thinking and solving
> problems;that it is not representational at all. That is, when we are
> coping at our best, ***we are drawn in by solicitations and respond directly
> to them, so that the distinction between us and our equipment--between inner
> and outer-vanishes***#1 As Heidegger sums it up:
> I live in the understanding of writing, illuminating, going-in-and-out, and
> the like. More precisely: as Dasein I am -- in speaking, going, and
> understanding -- an act of understanding dealing-with. My being in the
> world is nothing other than this already-operating-with-understanding in
> this mode of being.[ii]
>
A phantom limb is the sensation that an amputated or missing limb
(even an organ, like the appendix) is still attached to the body and
is moving appropriately with other body parts.
http://en.wikipedia.org/wiki/Phantom_limb
[A384] ...Now I maintain that all the difficulties commonly found in
these questions, and by means of which, as dogmatic objections, men
seek to gain credit for a deeper insight into the nature of things
than any to which the ordinary understanding can properly lay claim,
rest on a mere delusion by which they hypostatise what exists merely
in thought, and take it as a real object existing, in the same
character, outside the thinking subject.
In other words, they regard extension, which is nothing but
appearance, as a property of outer things that subsists [A385] even
apart from our sensibility, and hold that motion is due to these
things and really occurs in and by itself, apart from our senses.
http://www.arts.cuhk.edu.hk/Philosophy/Kant/cpr/
-----------------------------------
[Matter as Telepresence]
What is telepresence? I agree with Kac's (1997) distinction between
virtual reality (VR) and telepresence: VR presents purely synthetic
sense-data lacking physical reality. Telepresence presents sense-data
that (1) claims to correspond to a remote physical reality and (2)
allows the remote user to perform a physical action and see the
results. The WWW has the potential to bring telepresence out of the
laboratory.
http://www.walkerart.org/gallery9/beyondinterface/goldberg_artist.html
If I am just learning to use a hammer the first time I must extend the
perception of the end of my arm out about a foot further. I must learn
to hit an object, nail, a foot further out than I would normally hit
things. This is more like the telepresence in our adjustments to our
outer sense...
--------------------------------
The implementation of a Tele-Presence Microscopy Facility allows a
user from a remote location to either observe and/or control state-of-
the-art instrumentation in a real time interactive mode. ...The vision
suggested in the call for proposals is to combine all of these over
the National Information Infrastructure in such a way that makes
distributed collaboration as useful to scientific experiments as being
in the same location.
http://www.amc.anl.gov/docs/anl/tpm/tpmexecsumm.html
> Heidegger and Merleau-Ponty's understanding of embedded embodied coping,
> then, is not that the mind is sometimes extended into the world but rather
> that all such problem solving is derivative, that in our most basic way of
> being, that is, as absorbed skillful copers, we are not minds at all but one
> with the world. Heidegger sticks to the phenomenon, when he makes the
> strange-sounding claim that, in its most basic way of being, "Dasein is its
> world existingly."[iii]
>
Is this like, phenomenology, you know the attempt to extend the
inescapable self and somehow use it as proof that the external world
exists for certain by the process of eliminating material objects from
language and replacing them with hypothetical propositions about
observers and experiences, committing us to the existence of a new
class of ontological object altogether: the sensibilia or sense-data
which can exist independently of experience, and thus refute the
sceptics strong arguments?
Isaac
"Neil W Rickert" <ricke...@cs.niu.edu> wrote in message
news:yi5Uk.8809$ZP4....@nlpi067.nbdc.sbc.com...
> "Isaac" <gro...@sonic.net> writes:
>
>>Neil, I sent it to you email listed.
>
> Thanks. Copy received, and appreciated.
>
>> I replaced the "+" with "." assuming
>>that was an anti-SPAM measure.
>
> That was a mistake, and probably gave you a bounce. The "+"
> was correct. But fortunately you included both addresses, so
> the one with the "+" got through. (If you had removed the "+"
> and what follows it up to the "@", that would have worked to.
> The "+string" is to allow me to sort it into a different mailbox.
> And yes, that different mailbox does get extra anti-spam treatment.
>
I figured there was a chance it could be legit so I did both. Glad it
reached you.
> It's a bit long, so it might be a day or two before I have a comment.
If time is at a premium you may want to just focus on the sections that I
post critiques on. I plan on doing it in sequential order with the paper so
that might be more efficient for you.
> But it is rather interesting, and more interesting than some of
> Dreyfus's previous articles on AI.
He is always trying to bridge classic philosophy with technology to make
broad conclusions about AI. I think this time he is trying to rely on
cognitive neuroscience research (esp. in re to chaos and neural networks) to
buttress his long standing positions. However, I think he over reaches
their because he does not understand anything about the technical issues,
which imo greatly weakens, if not eviscerates his conclusions.
>I will probably be disagreeing
> with Dreyfus
On an AI news group, this is what I expect, of course.
>and possibly with you too.
Yes, I am counting on this being the case. You have two birds to shoot down
possibly with completely different ammunition. I am always very
controversial so it will be very easy for you to disagree with both Dryfus
and my comments.
>But then I have my own
> unique way of looking at questions of cognition.
Vive la differance!
I look forward to your well reasoned feedback on my critiques that I post
that you find interesting to respond to .
Ariel B.-
>
Immortalist
A learned skill like particular musical styles on the piano require
that the fingers move faster than conscious planning can accomplish.
The bridge in this case might be complex patterns stored in the
cerebellum, initiated by simple memory data in the basal ganglia,
which is triggered by some partial sense datum.
I think I agree but there is as yet no known way to escape from
skepticism concerning certainty on the issue. If memory is just the
stimulation of the same neurons that the senses stimulate, and this
information is stored in a much compressed style which can initiate
cascades that lead to the patterns, solipsism continues not to be
ruled out. It is still possible that we are in a matrix style delusion
with wires hoked to our brains.
......In a sparse distributed network - memory is a type of
perception.....The act of remembering and the act of perceiving both
detect a pattern in a vary large choice of possible patterns....When
we remember we recreate the act of the original perception - that is
we relocate the pattern by a process similar to the one we used to
perceive the pattern originally.
http://www.kk.org/outofcontrol/ch2-d.html
The question is really how rusty 20th century philosophy, albeit with
a Kantian twist, can deal with neuroscience discoveries, of the ebb
and flow of experience producing parts of the brain.
> Ariel B.
>
> "Alpha" <omegazero2...@yahoo.com> wrote in message
>
> news:f8636636-6ee8-4c92...@k24g2000pri.googlegroups.com...
>
> On Nov 14, 9:46 pm, "Isaac" <gro...@sonic.net> wrote:
>
> > All,
>
> >> I have critiqued in great detail a recent write paper by Prof. Hubert
> >> Dreyfus entitled "Why Heideggerian AI Failed and how Fixing it would
> >> Require
> >> making it more Heideggerian" . I can email a copy of it to whom ever is
> >> interested.
>
> >Please send a copy to omegazero2...@yahoo.com Ariel. Thanks!
Isaac
"Immortalist" <reanima...@yahoo.com> wrote in message
news:fcc39fdd-9886-401f...@h23g2000prf.googlegroups.com...
I did not understand the significance of the Kant, VR, phantom-limb, etc.
quotes. Please clarify what you mean in concrete terms.
> If I am just learning to use a hammer the first time I must extend the
> perception of the end of my arm out about a foot further. I must learn
> to hit an object, nail, a foot further out than I would normally hit
> things. This is more like the telepresence in our adjustments to our
> outer sense...
OK, but how does this support or contradict Dryfus' contention that we have
no internal representations that Dryfus quotes Heidegger as saying we have
in the above section?
>
>> Heidegger and Merleau-Ponty's understanding of embedded embodied coping,
>> then, is not that the mind is sometimes extended into the world but
>> rather
>> that all such problem solving is derivative, that in our most basic way
>> of
>> being, that is, as absorbed skillful copers, we are not minds at all but
>> one
>> with the world. Heidegger sticks to the phenomenon, when he makes the
>> strange-sounding claim that, in its most basic way of being, "Dasein is
>> its
>> world existingly."[iii]
>>
>
> Is this like, phenomenology, you know the attempt to extend the
> inescapable self and somehow use it as proof that the external world
> exists for certain by the process of eliminating material objects from
> language and replacing them with hypothetical propositions about
> observers and experiences, committing us to the existence of a new
> class of ontological object altogether: the sensibilia or sense-data
> which can exist independently of experience, and thus refute the
> sceptics strong arguments?
I think phenomenology as an objectified (i.e., represented) concept is what
Heidegger claims, but I understand Dryfus as saying here that he agrees with
Merleau-Ponty's philosophy that the phenominon and our perceiving of it
become indestingishable from each other; i.e, we have no internal
representations of the phenomenon.
Do you have any comments on my cretique of these paragraphs below?
<snip>
Immortalist
> "Immortalist" <reanimater_2...@yahoo.com> wrote in message
I suppose further conversation about the implications of telepresence
would and representationalism have not been ruled out by what Dyfus
says, unless he only wants to make a stronger but still probable
argument he shoots himself in the foot. Jeez, representationalism, is
the best theory since Kant and still rules neuroscience,
What is Representationalism?
Representationalism is the philosophical position that the world we
see in conscious experience is not the real world itself, but merely a
miniature virtual-reality replica of that world in an internal
representation. Representationalism is also known (in psychology) as
Indirect Perception, and (in philosophy) as Indirect Realism, or
Epistemological Dualism.
Why Representationalism?
As incredible as it might seem intuitively, representationalism is the
only alternative that is consistent with the facts of perception.
The Epistemological Fact (strongest theory): It is impossible to have
experience beyond the sensory surface.
Dreams, Hallucinations, and Visual Illusions clearly indicate that the
world of experience is not the same thing as the world itself.
The observed Properties of Phenomenal Perspective clearly indicate
that the world of experience is not the same as the external world
that it represents.
http://cns-alumni.bu.edu/~slehar/Representationalism.html
Representationalism (or indirect realism) with respect to perception
is the view that "we are never aware of physical objects, [but rather]
we are only indirectly aware of them, in virtue of a direct awareness
of an intermediary [mental] object. (Dancy, 145) Because there are
both direct and indirect objects of awareness in representationalism,
a correspondence relation arises between the mental entities directly
perceived and external objects which those mental entities represent.
And thus perceptual error occurs when the two objects of awareness do
not correspond sufficiently well. In opposition to
representationalism, both (direct) realism and idealism agree that
perception is direct and unmediated, despite their disagreements about
what the object of perception is. (Dancy, 145) In any form of direct
perception, no correspondence relationship is possible, since there is
only one object of perception. Thus only representationalism will give
rise to the view that perceptual errors exist and must be part of a
theory of perception. Nevertheless, both idealism and realism must
still account for the facts that are referred to as "perceptual
errors" by the representationalist.
http://www.dianahsieh.com/undergrad/rape.html
...representation is central to psychology as well, for the mind too
is a system that represents the world and possible worlds in various
ways. Our hopes, fears, beliefs, memories, perceptions, intentions,
and desires all involve our ideas about (our mental models of) the
world and other worlds. This is what humanist philosophers and
psychologists have always said, of course, but until recently they had
no support from science...
http://www.kurzweilai.net/meme/frame.html?main=/articles/art0162.html?
Plenty but your moving to fast, this could take months, even years,
with some of you hard cases.
Isaac
"Curt Welch" <cu...@kcwc.com> wrote in message
news:20081116234549.142$i...@newsreader.com...
> "Isaac" <gro...@sonic.net> wrote:
>> Reminder: I will post the paragraph(s) I have a comment about, and
>> highlight
>> the
>> particular words at issue by enclosing them between "***" characters.
>> I'll also include citations in the paper when helpful. I seek
>> (intelligent and informed) technical/theoretical critique or feedback
>> you are interested in the context and details.
>>
>> 2nd critique, on his page 12, line 4:
>> "Heidegger's important insight is not that, when we solve problems, we
>> sometimes make use of representational equipment outside our bodies, but
>> that being-in-the-world is more basic than thinking and solving
>> problems;that it is not representational at all. That is, when we are
>> coping at our best, ***we are drawn in by solicitations and respond
>> directly to them, so that the distinction between us and our
>> equipment--between inner and outer-vanishes***#1 As Heidegger sums it
>> up: I live in the understanding of writing, illuminating,
>> going-in-and-out, and the like. More precisely: as Dasein I am -- in
>> speaking, going, and understanding -- an act of understanding
>> dealing-with. My being in the world is nothing other than this
>> already-operating-with-understanding in this mode of being.[ii]
>>
>> then, is not that the mind is sometimes extended into the world but
>> rather that all such problem solving is derivative, that in our most
>> basic way of being, that is, as absorbed skillful copers, we are not
>> minds at all but one with the world. Heidegger sticks to the
>> phenomenon, when he makes the strange-sounding claim that, in its most
>> basic way of being, "Dasein is its world existingly."[iii]
>>
>> but, rather, that most basically we are absorbed copers, the inner/outer
>> distinction becomes problematic. There's no easily askable question as to
>> whether the absorbed coping is in me or in the world. According to
>> Heidegger, intentional content isn't in the mind, nor in some 3rd realm
>> (as it is for Husserl), nor in the world; it isn't anywhere. It's an
>> embodied way of being-towards. Thus for a Heideggerian, all forms of
>> cognitivist externalism presuppose a more basic existential externalism
>> where even to speak of "externalism" is misleading since such talk
>> presupposes a contrast with the internal. Compared to this genuinely
>> Heideggerian view, ***extended-mind externalism is contrived, trivial,
>> and irrelevant***#2.
>>
>>
<sniped citations>
>>
>> -------------------------------------------------------------------------
> I'm an engineer, not a philosopher.
>As such, nearly everything you write
> strikes me as silly and odd and misguided.
I am an engineer, scientist, philosopher, and roboticist. Of course, the
problem does not reside strictly in any one dicipline or skill set, so I am
not surprised that an implementation oriented thinker will find the
abstractions too obtuse for utility.
>I hardly know where to begin to
> comment.
>
> I find this sort of philosophical debate to be a pointless and endless
> game
> at trying to define, and redefine words to make them fit together in a
> more
> pleasing way. You can't solve AI by playing with words. You have to do
> it
> using empirical evidence.
I disagree. Reverse engineering will not solve the problem and may actually
lead to many dead ends. It will take a new theory and philosophy to do it.
Think of it like trying to emperically come up with QED or Relativity w/o
any new theory or philosophy of physics.
> It's not a problem which can be solved by pure
> philosophy.
>
True, but you can't just do it bottom up either. You can miss the big
picture, which philosophy can shed light on.
> For example, you speak of this "unity between the mind and the world".
> What exactly is the "mind" and the "world"?
I did not say this. If you read my intro, I was quoting from Dryfus' paper.
>You can't resolve this sort of
> question just by talking about such things. Words are defined by their
> connection to empirical evidence and without empirical evidence, the words
> are basically meaningless - or at minimal, available for use in endless
> pointless debates and redefinition based on usage alone.
for sure symantics can lead to circular definitions, but tossing out
anything not empiracle is "throughing the baby out with the bath water";
that is, you toss out powerful abstractions that bridge large gaps empirical
evidence.
>
> The problem we run into here is that without a concrete definition of how
> the brain works and what the mind is, we can't make any real progress on
> the types of issues you are touching on here. How can we make any
> progress
> debating the nature of the connection between the "mind" and the "world"
> when we can't agree what the mind is? And if we can't agree what the mind
> is, we can't really agree on anything it creates - like it's view of the
> world - which is the foundation of what the word "world" is referring to.
>
> You can't resolve any of these questions until you can first resolve
> fundamental questions such as the mind body problem and consciousness in
> general.
>
Well, we have to talk about the trinity or we'd get no where, but I agree
that any usage of those words must be very tentative and cannot lead to
sweeping conclusions w/o a scientific definition of each, which I say would
require a theory of mind (not connecting a million data points).
> I have my answers to these questions, but my answers are not shared or
> agreed on by any sort of majority of society so my foundational beliefs
> can't be used as any sort of proof of what is right. It call comes back
> to
> the requirement that we produce empirical data to back up our beliefs.
> And
> for this subject, that means we have to solve AI, and solve all the
> mysteries of the human brain. Once we have that hard empirical science
> work finished, then we will have the knowledge needed, to resolve the sort
> of philosophical debates you bring up here. Until then, endless debate
> about what "merging mind and world" might mean, is nearly pointless in my
> view.
>
> Having said all that, I'll give you my view of all this, and the answers
> to
> your questions as best as I can figure out.
>
> I'm a strict materialist or physicalist. I believe the brain is doing
> nothing more than performing a fairly straight forward signal processing
> function which is mapping sensory input data flows into effector output
> data flows. There is nothing else there happening, and nothing else that
> needs to be explained in terms of what the "mind" is or what
> "consciousness" is.
I don't think you can call anything as chaotic as the brain doing anything
"straight forward". The Earth's weather is infinitely more straitforward
than the humand mind/brain and we cannot model it worth a damn even with all
the most powerful computers in the world.
>The mind and consciousness is not something separate
> from the brain, it simply is the brain and what the brain is doing.
>
> It's often suggested that humans have a property of "consciousness" which
> doesn't exist in computers or maybe insects (based on the use of "insect"
> above). I see that idea as totally unsupported by the facts. It's nothing
> more than a popular myth - and a perfect example of the nonsense that is
> constantly batted around in these mostly pointless physiological debates.
>
> There is no major function which exists in the human brain which doesn't
> already exists in our computers and our robots which are already acting as
> autonomous agents interacting with their environment. The only difference
> between humans and robots, is that humans currently have a more advanced
> signal processing system - not one which is substantially different in any
> significant way - just one which is better mostly by measures of degree,
> and not measures of kind.
I don't think you could be farther away from the truth. The brain computes
in ways that is so different (an often oposite) of how our signal processing
works that it is in another universe by comparison. For example, the core
of the brain's sensory processing seems to be a kind of synethstesia based
system, which is exactly what all engineers would avoid like the plague. I
could go on and on with counter examples.
>
> Many people however tend to believe the human "mind" and human
> "consciousness" is something different from what our robots are doing by a
> major and important degree of kind. They believe we are something no one
> yet understands, and something that doesn't exist in our machines at all.
> I reject that notion completely.
>
> This belief we find in so many humans - that they are uniquely different
> from the machines - is a result of an invalid self-image the brains
> naturally tend to form about themselves. Human tend to think they are
> something they are not in this regard. They believe their "mind" is
> somehow different and separate from the brain, when there is no separation
> at all. The endless mind body debates and all the other debates which
> spin
> off from it, are the result of failing to see that the apparent separation
> is only an illusion.
>
> This illusion is represented by the simple idea that our internal
> awareness, doesn't seem to be an identity with neural activity. That
> "seeing blue" doesn't seem to us to be "just neural activity". Seeing
> blue
> seems to be something of a completely different nature to us than "neurons
> firing". However, all evidence adds up to the fact that these are an
> identity - that they are in fact one and the same thing. Not something
> "created by" the activity of the brain, but simply, the brain activity
> itself.
>
> Now, I wrote all that just to try and make it clear to you where I'm
> coming
> from and what I believe in.
>
> Because of what I believe, the mind body problem, and AI, and
> consciousness, translates to a very straight forward problem of science
> and
> engineering, not a philosophy problem in any respect. The brain is just a
> reinforcement trained parallel signal processing network which produces
> our
> behavior (both external behaviors and internal thoughts) as a reaction to
> the current environment.
hebbian learning was known since the '50's but that has not lead to anything
practical because it may necessary but not sufficient. For example, hebbian
learning does not even begin to solve the frame problem. Since this is so
strait forward, how do you propose reinforcement training (i.e., Pavlov's
dog) can be used to robustly deal with the frame problem?
>
> From this perceptive, let me jump in and debate the words you had issue
> with:
>
> we are drawn in by solicitations and respond
> directly to them, so that the distinction between us and our
> equipment--between inner and outer-vanishes
>
> I think at the lowest levels of what is happening in the brain, it is
> obvious that the brain is simply reacting to what is happening in the
> environment - that is what the brain is doing by definition in my view.
> We
> simply "respond directly to them". That is all we every do.
really? So, being an engineer you will know that "reactions" to input is
just another way of saying that you have a control system. However, any
control system needs a model to determine the proper control surface for the
input landscape; that is, model building. Thus, the brian is about building
useful models of the environment via sensory synergy. In this way, I
completely disagree with your assertion that "brain is simply reacting to
what is happening in the environment "
>
> But this is where it gets very messy. What is meant by "we" in the above?
>
> For me, it is obvious the only "we" that exists is a human body and the
> human body is simply reacting to its environment and that's pretty much
> the
> end of the story. It's no more complex or mystical in any sense than a
> robot reacting to its environment or a rock reacting to its environment.
> The only difference is that the more complex machines like the human and
> robot react in more complex ways to their environment than the rock does.
>
<sniped for brevity>
> If the part of the brain which represents "dogs" is damaged, we can become
> unable to see a dog.
Well, Dreyfus disagrees with you on this point. He says there is no
representation of a dog in the brain. How do you argue against that?
>We can look right at it, and have no clue what it is
> we are looking at.
All the evidence I am aware of re the brain is that such concepts are not
located in any one place which you can damage to lose only the recognition
of a dog. BTW, this is another example of how the brain is radically
different than our computing systems.
>At the same time, if you stimulate the correct parts of
> a brain, it's mostly likely that the person would report they were "seeing
> a dog" when there was no dog there.
There is no research ever showing that this is possible. Please cite the
research that supports your belief. I only know of music being able to be
stimulated to be heard in the brain.
>So what is it we are actually
> "seeing". Is it the dog we are responding to when we say we see a dog, or
> is the neural activity in one part of the brain which other part of the
> brain is responding to by producing the words "I see a dog"?
>
I think we should stay away from consciousness in this discussion or else we
will get no where by forking out to too many infinities.
> It can be argued that what we actually respond to is not the physical dog
> out in the word, but that we are responding to the brain activity.
Of course, the model of the dog.
> However, in a normally functioning human brain, the brain activity only
> shows up in the brain, when there is a dog in our field of vision - so the
> distinction isn't important.
>
> So when we use the the word "we", or "self" what actually are we talking
> about? Just like with the dog, are we talking about the phsyical thing
> "out there"? Or are we talking about the brain activity which represents
> the physical thing out there? The confusion here of course is that,
> unlike
> with the dog example, the physical thing out there and the brain activity,
> overlap. The brain activity is part of the human body which is thinking
> about itself.
Again, I think we should stay away from consciousness in this discussion or
else we will get no where by forking out to too many infinities. Starting
from lower levels of the brain is more practical here.
> Like when we use the word "dog" it's not clear whether we are talking
> about
> the brain activity which represents the dog, or the physical dog, it's not
> clear what we are talking about when we talk about "we".
>
> So lets go back to the words:
>
> we are drawn in by solicitations and respond
> directly to them, so that the distinction between us and our
> equipment--between inner and outer-vanishes
>
> Which "we" might these words be making reference to? I really don't know
> because I don't know for sure what this guy was trying to communicate with
> these words.
>
I believe he means our brain circuits engage phenominon by melding with it
and becoming a mirror image such that the two are not seprable, thus no
representations of the object in the brain just a bunch of organically
melded dominoes that hit one to another like "reality" would.
> But then we get to the "us" and "our equipment". Again, was the "us" and
> "our equipment" a reference to the physical body of the person and the
> physical equipment, or a reference to the brain activity inside the
> person's brain which represented all this stuff?
>
> And which "distinction" is he making reference to? The distinction
> between
> the human body and the physical stuff of the equipment? Or the
> distinction
> between the brain activity which represents the equipment and the brain
> activity which represents the human?
I believe he is saying that phenomenon is internalized w/o distinctions;
i.e., you become the phenomenon. As opossed to you making a model of the
object as a seeprate token to use in your brain system to plan your actions.
>
<snipped for brevity>
> As such, when the "dog" response is inhibited because some other response
> like "big scary lion which is about to eat us" is produced, we end up not
> "seeing" the dog. When asked later if there was a dog next to the lion,
> we
> will have no memory of the dog at all simply because we never "saw" the
> dog
> - i.e., the dog neural activity was blocked by other activity of higher
> priority. We would say, "we didn't notice the dog".
>
> This is true about most of how we respond to the environment. If we see a
> complex scene, the brain picks the reaction which it believes is the
> "best"
> for that situation - we will see one object first, and ignore everything
> else. We will "focus on" the object we "care about" the most.
>
I think you digress here. The issue is about cognitive architecture wrt to
phenomenon and building "correct" actions, not about focus of attention even
if that does act as an initial filter of what info we get as input to the
process. Again, let's stay away from conciousness here. That is a seperate
thread all together.
>
> This happens because when we are in a room, there will be all sorts of
> brain activity representing that room, and our location in the room.
> There
> is always brain activity which represents the state of the environment
> around us. But when we watch a TV show, or a movie, and we focus on the
> what is happening in the show long enough, the brain state which
> represents
> the fact that we are in a room watching a TV will start to fade and be
> slowly be replaced by the events of the show. The less distraction there
> is
> from the room (dark lights with no other noise or motion in the room) the
> greater this effect is.
>
> This is one effect that might be what the words "distinction between us
> and
> our equipment--between inner and outer-vanishes", could have been making
> reference to.
Dryfus is not talking about conciousness, he is talking about the
architecture of low level brain circuits that do ro do not make distinctions
between "us" and phenomenon.
> But then there's the other way to look at this. Just like the "dog"
> exists
> for us as brain activity, our entire world exists for us as brain
> activity.
> As such, there never really was much distinction between the brain
> activity
> which represents our inner self and brain activity which represents outer
> stuff. It's all still just "more brain activity". The ability to
> classify
> some brain activity as "outer stuff" and some brain activity as "inner
> stuff" is mostly a learned response. It's the way our brain is trained to
> respond to the enviro0nment - by creating that inner and outer
> classifications. As such, we can also train ourselves to think in terms
> of
> being "one with the world". Or "extend our consciousness outside
> ourselves" - or whatever sill way some one might talk about this.
>
In his paper, he considers chaotic neural networks as being more at "one
with the world" than classic AI's more (fuzzy) rule based systems.
<snipped>
> fact. Stop the brain activity, and my "world" goes away. I will no
> longer
> be "aware" of what is around me. I will no longer be "seeing" the
> computer
Again, let's stay away from awareness here. A seperate issue.
<snipped for brevity>
> I've not tried to comment on your critiques of the words yet, but this
> post
> is too long already so I'll stop hear for now.
>
Your detailed thougths are greatly appreciated. See if my feedback can help
focus ideas for or against Dreyfus' thesis or my critique of his ideas.
<snip>
> you bring up here must be translated back to the actions of such a machine
> in order to be understood. If you can't translates these ideas and
> questions and issues back to the operation of a machine like this, then
> you
> don't have any hope of understand what it is you are talking about and you
> don't have any hope of understanding what the words mean or why one set of
> words is a better or worse description of what is happening.
yes, but we do have to proceed with best working theories to give tentative
meaning to the symantics so we can work towards articulating the
distinctions we feel with scientific definititions... years in the future.
Best regards,
Ariel B.
>
> --
> Curt Welch
> http://CurtWelch.Com/
> cu...@kcwc.com
> http://NewsReader.Com/
Josip Almasi
...
> [i] As Heidegger puts it: "The self must forget itself if, lost in the world
> of equipment, it is to be able 'actually' to go to work and manipulate
> something." Being and Time, 405.
> My critique #1:
>
> seems that the "distinction between us and our equipment... is vanished" is
> just describing the unconscious automation process that takes over body
> functions and relieves the conscious mind to be unaware that its equipment
> was drawn into responding to solicitations. This in many ways seems to just
> be alluding to the domain of our unconscious being that responds like
> dominos that fall automatically in response to many contextual
> solicitations.
I think it's simply about ego.
Lack of ego doesn't mean being unconscious.
It's just an observation switch, seing how things relate to each other,
rather than how relate to self. IOW, 'how to' instead of 'how do I'.
> I do not see how this all makes a solid argument that
> conscious thought is unified and inseparable from "our equipment" (i.e.,
> body). At best this is a very weak, if not completely flawed, logic in
> inferring that our sense (act) of being in the world "is not
> representational at all".
Agreed.
> This all seems to be logically
> flawed and/or a very weak foundation for grandiose conclusions about what
> philosophical approach/model is needed to solve the frame problem and human
> consciousness. Maybe I am missing something critical here that can make
> sense of it. Please clarify the logic.
Agreed.
Regards...
Neil W Rickert
>I disagree. Reverse engineering will not solve the problem and may actually
>lead to many dead ends. It will take a new theory and philosophy to do it.
>Think of it like trying to emperically come up with QED or Relativity w/o
>any new theory or philosophy of physics.
It's good to see someone who recognizes that reverse engineering
is not sufficient. To few people recognize this.
>I did not say this. If you read my intro, I was quoting from Dryfus' paper.
Quick comment. You are frequently writing "Dryfus" instead of "Dreyfus".
>I don't think you can call anything as chaotic as the brain doing anything
>"straight forward". The Earth's weather is infinitely more straitforward
>than the humand mind/brain and we cannot model it worth a damn even with all
>the most powerful computers in the world.
I doubt that the brain is chaotic, except perhaps during epileptic
seizures and similar failures. Perhaps you are using "chaotic" only
to mean that we don't have a satisfactory theory of brain operations.
>> There is no major function which exists in the human brain which doesn't
>> already exists in our computers and our robots which are already acting as
>> autonomous agents interacting with their environment. The only difference
>> between humans and robots, is that humans currently have a more advanced
>> signal processing system - not one which is substantially different in any
>> significant way - just one which is better mostly by measures of degree,
>> and not measures of kind.
>I don't think you could be farther away from the truth. The brain computes
>in ways that is so different (an often oposite) of how our signal processing
>works that it is in another universe by comparison. For example, the core
>of the brain's sensory processing seems to be a kind of synethstesia based
>system, which is exactly what all engineers would avoid like the plague. I
>could go on and on with counter examples.
Part of the confusion is in the assumption that the brain is computing.
I find little evidence of that. By the way, I have had these debates
with Curt in the past.
>hebbian learning was known since the '50's but that has not lead to anything
>practical because it may necessary but not sufficient. For example, hebbian
>learning does not even begin to solve the frame problem. Since this is so
>strait forward, how do you propose reinforcement training (i.e., Pavlov's
>dog) can be used to robustly deal with the frame problem?
Perhaps Hebbian learning is not properly understood. At least that's
part of my position.
IMO there is no need to solve the frame problem. Rather, we
need to avoid it. The frame problem is simply an artifact of the
reliance on stored representations. Humans suffer from the frame
problem when they depend on stored representations. I remember
some time back when I replied to a want ad in the newspaper, and
left a message on the person's tape. I was called back twice -
apparently he had forgotten that he had called me back the first
time (i.e. he had failed to update his stored representations).
And that seems to be an example of a frame problem failure.
>All the evidence I am aware of re the brain is that such concepts are not
>located in any one place which you can damage to lose only the recognition
>of a dog. BTW, this is another example of how the brain is radically
>different than our computing systems.
Not a good example. Back in the early days of the PC, a single byte
of memory was actually distributed over 9 different chips plugged
into 9 different sockets on the PC. That one byte of memory was all
in one location according to our formal models of computing, but was
distributed over multiple components in the actual implementation.
Incidently, I do agree that brains are radically different from computing
systems. But that's a theoretical view, not something that can easily
be determined empirically.
Neil W Rickert
>What is Representationalism?
>Representationalism is the philosophical position that the world we
>see in conscious experience is not the real world itself, but merely a
>miniature virtual-reality replica of that world in an internal
>representation. Representationalism is also known (in psychology) as
>Indirect Perception, and (in philosophy) as Indirect Realism, or
>Epistemological Dualism.
>Why Representationalism?
>As incredible as it might seem intuitively, representationalism is the
>only alternative that is consistent with the facts of perception.
Nonsense.
And yes, I have had this debate with Steven Lehar. Neither of us
was able to persuade the other.
I would suggest that J.J. Gibson's direct perception is more
plausible than Lehar's version of representationalism.
Isaac
a very rough summary of Dreyfus' thesis in this paper:
In summ: Dryfus's paper is not really talking about consciousness, he is
talking about the architecture of low level brain circuits that do or do not
make distinctions (e.g., representations) between "us" and phenomenon. He
roughly says our brain circuits engage phenomenon by melding with it and
becoming a mirror image such that the two are not separable, thus no
representations of the object or modules or hierarchy in the brain (which he
asserts is what brings the downfall of Brooks, Minsky, and the like) just a
bunch of flat, organically melded dominoes that hit one to another like
"reality" would. He says that the "representation" approach/architecture is
of Heidegger's philosophy, and hence forth Heideggerian AI, and the
non-representation approach/architecture is espoused by Merleau-Ponty,
Walter Freeman. He particularly hangs his hat of Walter Freeman's
neurodynamic model as the solution to AI; i.e., a chaotic, flat, neural
network approach.
Here are a couple good quotes to summ up his Thesis:
"... there are now at least three versions of supposedly Heideggerian AI
that might be thought of as articulating a new paradigm for the field:
Rodney Brooks' behaviorist approach at MIT, Phil Agre's pragmatist model,
and Walter Freeman's neurodynamic model. All three approaches implicitly
accept Heidegger's critique of Cartesian internalist representations, and,
embrace John Haugeland's slogan that cognition is embedded and embodied.[i]
.... Later I'll suggest that Walter Freeman's neurodynamics offers a
radically new basis for a Heideggerian approach to human intelligence-an
approach compatible with physics and grounded in the neuroscience of
perception and action. But first we need to examine another approach to AI
contemporaneous with Brooks' that actually calls itself Heideggerian.
--------------------------------------------------------------------------------
[i] John Haugeland, "Mind Embodied and Embedded," Having Thought: Essays in
the Metaphysics of Mind, (Cambridge, MA: Harvard University Press, 1998),
218."
I hope this helps frame the issues and my critiques better for those who
understandably do not have time to read his paper.
Cheers,
Ariel B.
"Alpha" <omegaz...@yahoo.com> wrote in message
news:f8636636-6ee8-4c92...@k24g2000pri.googlegroups.com...
On Nov 14, 9:46 pm, "Isaac" <gro...@sonic.net> wrote:
> All,
>
> I have critiqued in great detail a recent write paper by Prof. Hubert
> Dreyfus entitled "Why Heideggerian AI Failed and how Fixing it would
> Require
> making it more Heideggerian" . I can email a copy of it to whom ever is
> interested.
Please send a copy to omegaz...@yahoo.com Ariel. Thanks!
Why don't you post a summary of that paper and your key critique
points here.
Isaac
<forbi...@msn.com> wrote in message
news:36705361-5c01-4d0f...@c36g2000prc.googlegroups.com...
> You can send me the paper.
Done. It should be in your inbox by now.
>I was impressed by
> Dreyfus decades ago. Since I find myself untrainable
> within the University environment I have many weak
> spots. This being said, I have participated in the
> c.a.p forum since its splt form c.a so as to move
> the Searle chatter out of c.a.
great, then you may be bent towards sticking up for him. Of course, most in
the AI community would not. So, any counter balance is very needed.
>
> While traditional Representationalism is unlikely
> to be true it seem just as unlikely thinking bridges
> any gaps. I think I agree with you on this point.
>
> I suspect we are like the blind men describing
> an elephant. We seem to argue over descriptions
> rather than trying to integrate them into a cohesive
> whole. I'm as guilty as the next.
>
Yes, that is why getting the philosophy and theory strait is an important
first step in the right direction.
> Most of us are much harder on papers when we
> disagree with their conclusions than when we
> agree with them.
I'm just as hard either way. I don't let anyone agree with me unless they
do it with the right rational, and can strongly defend it against my strong
ability to play the devil's advocate.
>Are you sure you came to the
> table clean?
I did. Actually, I tend to agree with some architectural implications of
his philosophy, however, I believe his reasoning is very flawed so I
critique it at every step.
>
> I tend to shut down as soon as I reach an unsupported
> claim with which I disagree. It doesn't really matter
> that much what conclusion is reached. This has really
> interfered with my reading of technical papers. On the
> other hand I can accept hypotheticals I strongly suspect
> are false.
I understand how that feels. I can stick with it if they give strong
reasoning and/or plausible evidence. I tune out when there is too much hand
waving.
>Is the paper presented as a series of conclusions
> one must accept as true or as alternatives to be considered?
No, he does do a reasonably good job at grounding it; however, when he does
make (many) broad conclusions, I believe he jumps too far. So, I'd give him
a C+ on this count, where most, if not nearly all, philosophers would give a
D- on a practicality rating scale.
Thanks for your input,
Ariel-
Isaac
> I'll begin my comments here.
>
> But first a general comment. AI is pretty much a mechanization of
> traditional epistemology. Dreyfus seems to acknowledge this in the
> first couple of pages. So if AI does not work, that would suggest
> a problem with epistemology. So why do people like Dreyfus (and
> Searle, and others) criticize AI but not criticize epistemology?
> (OK, that's a rhetorical question).
Dryfus' solution to this is to say AI's model of epistemology is wrong and
needs to be embodied, flat, and non-representational.
>
> To me, epistemology has always seemed a bit silly. And it has
> puzzled me that intelligent philosophers fail to see that it
> is silly.
>
Epistemology is not really the same scope as AI. Beyond learning and
building knowledge, AI also includes transcendental aspects of consciousness
and self (soul?), which are in metaphysics. AI also covers the creation and
appreciation of beautiful things, which is in the 3rd pillar of philosophy:
esthetics. So, I believe AI touches on nearly all aspects of philosophy.
> "Isaac" <gro...@sonic.net> writes:
>
>>First, critique, page 11, line 20:
>>"I agree that it is time for a positive account of Heideggerian AI and of
>>an
>>underlying Heideggerian neuroscience, but I think Wheeler is the one
>>looking
>>in the wrong place. Merely by supposing that Heidegger is concerned with
>>problem solving and action oriented representations, Wheeler's project
>>reflects not a step beyond Agre but a regression to aspects of pre-Brooks
>>GOFAI. Heidegger, indeed, claims that that skillful coping is basic, but
>>he
>>is also clear that, all coping takes place on the background coping he
>>calls
>>being-in-the-world that doesn't involve any form of representation at all.
>
> Comment (on Dreyfus): I have no problem with the idea of skillful
> coping. The trouble I have with some of the philosophy, is that
> it often tends to come across as mystical in its reliance on vague
> ideas such as "being in the world".
>
>>see: Michael Wheeler, Reconstructing the Cognitive World, 222-223.
>
>>My comment:
>>"Assuming that by "thinking" you mean conscious thought, I cannot see how
>>thinking is a bridge that necessarily follows from memories/beliefs not
>>being solely inner entities. It seems to me that inner and outer
>>representations can be bridged without thought. Isn't this what occurs in
>>an unconscious (reflex) reaction to a complex external even, which is an
>>automatic bridge and generates a thoughtful, usually accurate response but
>>often before we even have a chance to think about it. Inner/outer
>>representations seems semantically vague here. Also, cannot conscious
>>thought can endeavor itself with in purely inner or out representations
>>without ever bridging them? I guess, it is the "therefore" that gives me
>>pause here."
>
> In a way, you are saying something similar to what I said above.
> That is, you are decrying the tendency to give accounts that seem
> mystical because of their reliance on rather vague ideas.
>
he does ground it more later in the paper, which I will present and comment
on little by little. However, imho, I think is pattern of making huge leaps
after "therefore" stays the same.
Cheers,
Ariel-
Isaac
AI also covers the creation and appreciation of beautiful things, which is
in the 3rd pillar of philosophy: esthetics. So, I believe AI touches on
nearly all aspects of philosophy. Moreover, (reverse) engineering will not
solve the problem and may actually lead to many dead ends by just finding
ways to go nowhere quicker and better. It will take a new theory and
philosophy to do it.
Think of it like trying to empirically come up with QED or Relativity w/o
any new theory or philosophy of physics.
"ŠućMućPaProlij" <123...@654321.00> wrote in message
news:gfs0lv$bvq$1...@ss408.t-com.hr...
> >
>> I'm an engineer, not a philosopher. As such, nearly everything you write
>> strikes me as silly and odd and misguided. I hardly know where to begin
>> to
>> comment.
>>
>> I find this sort of philosophical debate to be a pointless and endless
>> game
>> at trying to define, and redefine words to make them fit together in a
>> more
>> pleasing way. You can't solve AI by playing with words. You have to do
>> it
>> using empirical evidence. It's not a problem which can be solved by pure
>> philosophy.
>>
>
> I agree with you. Creating AI has nothing to do with philosophy. It is
> just a technical problem that needs better mathematical tools in order to
> solve it.
>
> Creating AI will reflect on philosophy in only one way - it will prove
> that some philosophers were wrong.
>
Isaac
model as the solution to AI; i.e., a chaotic, flat, neural network approach,
which he contends in this paper has no representations, modules, or
hierarchy.
"Immortalist" <reanima...@yahoo.com> wrote in message
news:66f01963-fb3e-41be...@d10g2000pra.googlegroups.com...
Isaac
thanks,
Ariel-
Isaac
"Josip Almasi" <j...@vrspace.org> wrote in message
news:gfrpua$jbm$1...@news.metronet.hr...
> Isaac wrote:
> ...
>> [i] As Heidegger puts it: "The self must forget itself if, lost in the
>> world
>> of equipment, it is to be able 'actually' to go to work and manipulate
>> something." Being and Time, 405.
> ...
>> My critique #1:
>>
>> seems that the "distinction between us and our equipment... is vanished"
>> is
>> just describing the unconscious automation process that takes over body
>> functions and relieves the conscious mind to be unaware that its
>> equipment
>> was drawn into responding to solicitations. This in many ways seems to
>> just
>> be alluding to the domain of our unconscious being that responds like
>> dominos that fall automatically in response to many contextual
>> solicitations.
>
> I think it's simply about ego.
> Lack of ego doesn't mean being unconscious.
> It's just an observation switch, seing how things relate to each other,
> rather than how relate to self. IOW, 'how to' instead of 'how do I'.
>
that is an intersting idea for matters that require a self-centered account
of phenomenon; however, how does your "relational ego" address the issue
raised about Dryfus' assertion that our brain/mind fundamentally makes no
"distinction between us and our equipment" so that we are simply, and
automatically, drawn into responding to solicitations like water flowing
down a hill?
Isaac
> "Isaac" <gro...@sonic.net> writes:
>>"Curt Welch" <cu...@kcwc.com> wrote in message
>
>>I disagree. Reverse engineering will not solve the problem and may
>>actually
>>lead to many dead ends. It will take a new theory and philosophy to do
>>it.
>>Think of it like trying to emperically come up with QED or Relativity w/o
>>any new theory or philosophy of physics.
>
> It's good to see someone who recognizes that reverse engineering
> is not sufficient. To few people recognize this.
>
>>I did not say this. If you read my intro, I was quoting from Dryfus'
>>paper.
>
> Quick comment. You are frequently writing "Dryfus" instead of "Dreyfus".
thanks,
>
>>I don't think you can call anything as chaotic as the brain doing anything
>>"straight forward". The Earth's weather is infinitely more straitforward
>>than the humand mind/brain and we cannot model it worth a damn even with
>>all
>>the most powerful computers in the world.
>
> I doubt that the brain is chaotic, except perhaps during epileptic
> seizures and similar failures. Perhaps you are using "chaotic" only
> to mean that we don't have a satisfactory theory of brain operations.
>
Dreyfus' solution to the "failed AI" is an embodied system based on a
chaotic neural network like that of Walter Freeman's neurodynamics. So, how
do you argue against this hypothesis?
<snip>
>
>>I don't think you could be farther away from the truth. The brain
>>computes
>>in ways that is so different (an often oposite) of how our signal
>>processing
>>works that it is in another universe by comparison. For example, the core
>>of the brain's sensory processing seems to be a kind of synethstesia based
>>system, which is exactly what all engineers would avoid like the plague.
>>I
>>could go on and on with counter examples.
>
> Part of the confusion is in the assumption that the brain is computing.
> I find little evidence of that. By the way, I have had these debates
> with Curt in the past.
>
Interesting, how do you define "computation" such that the brain is not
doing it, but our computers are? For example, when our brain perceives an
object and generates a motor plan to grab the object, is the brain not, even
if implicitely, performing calculation based on past data to create a
solution to a complex "reality" landscape equation?
>>hebbian learning was known since the '50's but that has not lead to
>>anything
>>practical because it may necessary but not sufficient. For example,
>>hebbian
>>learning does not even begin to solve the frame problem. Since this is so
>>strait forward, how do you propose reinforcement training (i.e., Pavlov's
>>dog) can be used to robustly deal with the frame problem?
>
> Perhaps Hebbian learning is not properly understood. At least that's
> part of my position.
>
> IMO there is no need to solve the frame problem. Rather, we
> need to avoid it. The frame problem is simply an artifact of the
> reliance on stored representations. Humans suffer from the frame
> problem when they depend on stored representations. I remember
> some time back when I replied to a want ad in the newspaper, and
> left a message on the person's tape. I was called back twice -
> apparently he had forgotten that he had called me back the first
> time (i.e. he had failed to update his stored representations).
> And that seems to be an example of a frame problem failure.
>
Isn't the frame problem mostly (if not all) about filtering the intractable
sensory information of any situation into a Gestalt of only meaningful,
important information. This seems to be along the lines of "common sense",
which you cannot just ingore and expect to meet or exceed human intelligence
or behavior skills.
>>All the evidence I am aware of re the brain is that such concepts are not
>>located in any one place which you can damage to lose only the recognition
>>of a dog. BTW, this is another example of how the brain is radically
>>different than our computing systems.
>
> Not a good example. Back in the early days of the PC, a single byte
> of memory was actually distributed over 9 different chips plugged
> into 9 different sockets on the PC. That one byte of memory was all
> in one location according to our formal models of computing, but was
> distributed over multiple components in the actual implementation.
>
> Incidently, I do agree that brains are radically different from computing
> systems. But that's a theoretical view, not something that can easily
> be determined empirically.
if you have an analog computation with transistors, or an optical
transformation (i.e., calculation) with a Fresnel lens, isn't computation
always present when a system transforms an input to a more useful output
used by another part of the system? For example, our eye does so many
critical calculation to signal condition the optical stream. Are you saying
that our eyes do not do any calculations? Please clarify without making the
semantics more vague.
thanks,
Ariel-
ŠućMućPaProlij
>transcendental aspects of consciousness and self (soul?), which are in
>metaphysics. Do you really think there is an E=mC^2 equation for that?
>
I don't know what "transcendental aspects of consciousness and self (soul?)" are
and I don't care. I leave this to philosophers.
> AI also covers the creation and appreciation of beautiful things, which is in
> the 3rd pillar of philosophy: esthetics. So, I believe AI touches on nearly
> all aspects of philosophy. Moreover, (reverse) engineering will not solve the
> problem and may actually lead to many dead ends by just finding ways to go
> nowhere quicker and better. It will take a new theory and philosophy to do
> it.
> Think of it like trying to empirically come up with QED or Relativity w/o any
> new theory or philosophy of physics.
>
I can tell you one thing - if we must wait philosophers to tell us how to make
AI then philosophers who think it is impossible to make AI are right.
Curt Welch
> I completely disagree, Beyond learning and building knowledge, AI also
> includes transcendental aspects of consciousness and self (soul?), which
> are in metaphysics. Do you really think there is an E=mC^2 equation for
> that?
Most definitely. As I said, I'm a strict physicalist. To me, the belief
that consciousness is something other than physical brain function is just
a widely held illusion or myth. People believe it, and debate it, just
like they waste their time believing, and debating, the nature of God.
It's just silly crap man made up for reasons that have nothing to do with
the nature of reality.
> AI also covers the creation and appreciation of beautiful things, which
> is in the 3rd pillar of philosophy: esthetics.
Beauty is created by the value the brain assigns to sensations and those
values are there because humans are reinforcement learning machines.
There's nothing more to it than that. Beauty isn't a mystery. It's simple
and obvious once you understand what we are - reinforcement learning
machines.
However, this is exactly the type of thing which is nearly impossible to
understand by using philosophy alone to try and uncover the nature of
beauty.
> So, I believe AI touches
> on nearly all aspects of philosophy.
Yes, I agree completely with that. Philosophy is one of many human
behaviors and if you don't understand where human behavior comes from and
what controls it, you will have no hope of answering any of the big
questions of philosophy such as the mind body problem and the nature of
consciousness or the nature of aesthetics. Those questions can't be
answered from within the field of philosophy alone. All you can do from
within philosophy is identify which concepts are compatible with each other
and which are not - you can't identity which set of beliefs are a valid
description of reality without checking the beliefs against empirical data
- which is something philosophy chooses to treat as being outside their
domain.
All you can do from within philosophy is create multiple possible answers.
You can't tell which is correct or how correct or incorrect a given
approach might be.
> Moreover, (reverse) engineering
> will not solve the problem and may actually lead to many dead ends by
> just finding ways to go nowhere quicker and better. It will take a new
> theory and philosophy to do it.
Reverse engineering has already solved it. Many philosophers however don't
understand this because they have created such a huge cloud of confusion by
spending so much time debating all the impossible answers they can't get a
grip on what the truth is.
> Think of it like trying to empirically come up with QED or Relativity w/o
> any new theory or philosophy of physics.
You have started with the assumption that there is something there
(consciousness) which is fundamentally hard to understand and explain.
Your assumption is invalid. Your assumption is created by a simple to
explain brain function which created in all of us a natural illusion. If
you assume the illusion is real, you are left with the hard problem of
consciousness. If you assume the illusion is only an illusion, then there
is no problem at all - all hard questions are answered and explained
leaving a fairly simple material world to understand. By Occam's razor, I
choose the answer that makes everything simple and answers all the
questions instead of picking the answer which creates contractions that
have no answer.
But in philosophy, Occam's razor has no place. All alternatives must be
explored - as such, you are forced by your very charter to wander endlessly
into utter silliness. The hardness of the problem attracts you to explore
the depths of the silliness endlessly hoping to find some "new theory" or
sudden enlightened understanding which clears it all up.
As an engineer, I have no need to explore such an improbable dead end. If
I missed something, the philosophers of the world will find it and explain
it. But after hundreds of years not finding an answer, I'm not holding my
breath on the expectation that there is something there when all evidence
suggests there isn't anything there to be found.
forbi...@msn.com
and skimmed the rest. It's quite dense. I won't be able
to get a good feel for several days because I come home
quite tired.
Based upon the conculsion I wonder why Stevan
Harnad's Total Turing Test wasn't mentioned.
I remeber reading "What Computers Can't Do"
in the early '70s, maybe '72 when it came out.
I don't remember much of the book any more.
I should read it again. It set the stage for
my reading 1980 reading of John Searle's
article in Behavioral and Brain Sciences.
Still, He's come quite a ways to make the
statement:
We can, however, make some progress towards animal AI.
So many of us have our reasons to doubt computers
(as currently constructed and envisioned) will be able
to cheaply simulate situated human coping but each
has his or her own reason.
I am encouraged by this paragraph:
So, according to the view I have been presenting,
even if the Heideggerian/Merleau-Pontian approach
to AI suggested by Freeman is ontologically sound
in a way that GOFAI and subsequent supposedly
Heideggerian models proposed by Brooks, Agre, and
Wheeler are not, a neurodynamic computer model
would still have to be given a detailed description of
a body and motivations like ours if things were to
count as significant for it so that it could learn to act
intelligently in our world. We have seen that Heidegger,
Merleau-Ponty, and Freeman offer us hints of the elaborate
and subtle body and brain structures we would have to
model and how to model some of them, but this only
makes the task of a Heideggerian AI seem all the more
difficult and casts doubt on whether we will ever be able
to accomplish it.
Even if not intended he shows the way. Still, digital systems
in digital worlds are subject to latching where analog systems
in analog worlds would not. As long as the designed system
is representational, even if of the system being modelled, it
is GOFAI all the way down.
I believe your sentence parsing for the sentence starting page 11
line 33 is incorrect in that Dryfus is attributing a conclusion to
Chalmers, Clark, and Wheeler rather than making one himself.
See the next paragraph to see why I believe so.
Isaac
> Publius wrote:
>>
>> Disputes about representationalism appear in AI discussions because the
>> disputants are not distinguishing between intelligence and consciousness.
>> The latter almost certainly entails representationalism; the former need
>> not, but natural intelligent systems may employ it. It's an empirical
>> question.
>
> Actually there's yet another big question - is consciousness emergent
> property of (sufficiently high) intelligence.
> And there's an empirical part of it - how to make it:)
> But seems noone around doubts it is an emergent property, so no need to
> emphasise distinction.
I tend to disagree. I do not see the two being so intimately connected so
as to require one to immerge from another. Intelligence might just be a
boundary condition on the scope of consciousness (i.e., awareness). For
example, a severely retarded person is certainly far less intelligent,
however, I don't think there is any evidence that they are far less
conscious. If consciouness emerged (necessarily?) from intelligence then
shouldn't they be highly correlated?
> As for empirical part, representationalism is sort of top-down approach,
> while say Curt seems to prefer bottom-up approach.
> Can't say for Dreyfus though:)
Dreyfus' paper necessarily requires a bottom-up approach because his
architecture has no hierarchy, no modules, and no representations.
Best,
Ariel-
>
> Regards...
Isaac
"Publius" <m.pu...@nospam.comcast.net> wrote in message
news:Xns9B59A2F03AC17mp...@69.16.185.250...
> "Isaac" <gro...@sonic.net> wrote in
> news:49212a1c$0$33562$742e...@news.sonic.net:
>
<snip , comment noted>
> From another post of yours:
>
> "Dryfus says that the brain works according to Walter Freeman's
> neurodynamic model as the solution to AI; i.e., a chaotic, flat, neural
> network approach, which he contends in this paper has no representations,
> modules, or hierarchy."
>
> Disputes about representationalism appear in AI discussions because the
> disputants are not distinguishing between intelligence and consciousness.
> The latter almost certainly entails representationalism; the former need
> not, but natural intelligent systems may employ it. It's an empirical
> question.
I tend to disagree generally, but do agree that Dreyfus' arguments apply the
most to low level systems like the limbic system. This is one of my
critiques as well. However, your solution seems to not work for me either.
That is, any intelligent system must generate abstractions off of highly
detailed phenomenon experienced. Are not abstractions a generic
representation of the more complex phenomenon being modeled for use in
either reasoning or pattern recognition. To me, if you have no
representations in your system then you must use the full resolution of the
experience phenomenon, which would result in a very brittle system because
it lacks abstractions that can enable the intelligent system to manipulate
intractable details as a single, simple package. Please clarify your idea
in the context of my points above.
,
Ariel-
Neil W Rickert
>"Neil W Rickert" <ricke...@cs.niu.edu> wrote in message
>news:pchUk.6319$Ei5....@flpi143.ffdc.sbc.com...
>> I doubt that the brain is chaotic, except perhaps during epileptic
>> seizures and similar failures. Perhaps you are using "chaotic" only
>> to mean that we don't have a satisfactory theory of brain operations.
>Dreyfus' solution to the "failed AI" is an embodied system based on a
>chaotic neural network like that of Walter Freeman's neurodynamics. So, how
>do you argue against this hypothesis?
If it is chaotic, then one might reasonably expect human behavior
to be chaotic. But that seem a rather odd way of characterizing
human behavior.
>> Part of the confusion is in the assumption that the brain is computing.
>> I find little evidence of that. By the way, I have had these debates
>> with Curt in the past.
>Interesting, how do you define "computation" such that the brain is not
>doing it, but our computers are? For example, when our brain perceives an
>object and generates a motor plan to grab the object, is the brain not, even
>if implicitely, performing calculation based on past data to create a
>solution to a complex "reality" landscape equation?
I am skeptical that there is much being retained in the form of
stored representations as "past data". It seems more likely that
the brain is a bit like a finely tuned instrument. The past data
has played a role in adjusting the tuning, but has not been retained.
The "motor plan" to grab an object is likely quite crude, and
precise behavior results not from having a precise plan, but from
measuring performance during the motor action and adjusting it where
the measurement indicates it is off. This would make measurement
more important than computation.
>Isn't the frame problem mostly (if not all) about filtering the intractable
>sensory information of any situation into a Gestalt of only meaningful,
>important information. This seems to be along the lines of "common sense",
>which you cannot just ingore and expect to meet or exceed human intelligence
>or behavior skills.
That's not the way the frame problem is usually described.
See http://en.wikipedia.org/wiki/Frame_problem for a more familiar
version. In any case, what matters here is the version of the frame
problem being assumed by Dreyfus in the paper we are discussing.
And I'm pretty sure he is taking it as the problem of updating
stored representations so that they are consistent with changes in
the world. Thus he mentions that Rodney Brooks avoids the problem
by designing his system to not depend on stored representations of
the state of the world.
>if you have an analog computation with transistors, or an optical
>transformation (i.e., calculation) with a Fresnel lens, isn't computation
>always present when a system transforms an input to a more useful output
>used by another part of the system? For example, our eye does so many
>critical calculation to signal condition the optical stream. Are you saying
>that our eyes do not do any calculations? Please clarify without making the
>semantics more vague.
You seem to be arguing that when I beep the horn on my car, the
mechanism in the car is somehow performing a calculation. If you
take the meaning of "computation" to be that broad, then everything
is a computation, and the word "computation" becomes useless for
it fails to discriminate.
Neil W Rickert
>> I would suggest that J.J. Gibson's direct perception is more
>> plausible than Lehar's version of representationalism.
>Could you briefly summarize the Thesis of J.J. Gibson's direct perception?
Gibson's view is that the perceptual system contains transducers
that are tuned to the invariants of various objects, and that with
the use of these transducers the objects are directly recognized
(identified) without and intermediate representation and without
any computation. Presumably learning involves the construction
and tuning of the transducers.
When at the supermarket, there is a scanner that reads the bar codes
on the items I am buying. The way the bar code scanner works seems
to better fit Gibson's account than Lehar's representational account.
There are good pragmatic reasons why the bar code scanner works
that way, and the same kind of pragmatic considerations are likely
to have been in play in what made it through biological evolution.
casey
> "Josip Almasi" <j...@vrspace.org> wrote in message
> > Actually there's yet another big question - is consciousness emergent
> > property of (sufficiently high) intelligence.
> > And there's an empirical part of it - how to make it:)
> > But seems noone around doubts it is an emergent property, so no need to
> > emphasise distinction.
>
> I tend to disagree. I do not see the two being so intimately connected so
> as to require one to immerge from another. Intelligence might just be a
> boundary condition on the scope of consciousness (i.e., awareness). For
> example, a severely retarded person is certainly far less intelligent,
> however, I don't think there is any evidence that they are far less
> conscious. If consciouness emerged (necessarily?) from intelligence then
> shouldn't they be highly correlated?
The degree of awareness certainly doesn't seem ot correlate directly
with
the degree of intelligent behavior shown. Intelligent behavior I
suspect also
depends on how much the agent is aware of and the quality of the
things
the agent is aware of.
JC
turtoni
"good" questions and perhaps create as a by-product some kind of
artificial "intelligence". After all that would appear to be how
"intelligence" got started.
So for example; how do we create an *actual* unlimited supply of
renewable energy?
I imagine that from these types of problems a form of artificial
intelligence will arise.
Curt Welch
> I think we would be more successful in continuing to work at some
> "good" questions and perhaps create as a by-product some kind of
> artificial "intelligence". After all that would appear to be how
> "intelligence" got started.
>
> So for example; how do we create an *actual* unlimited supply of
> renewable energy?
Well, that's a silly question. There's no such thing as "renewable
energy".
The best energy source around here is the sun, so the real energy problem
is simply a question of how do we best tap the energy flow from the sun.
Answering that gives as a few billion years of energy (not renewable).
> I imagine that from these types of problems a form of artificial
> intelligence will arise.
Well, the better question along that line is more like:
How do we build a machine that can survive on its own without human help?
People working on real world robotics questions are already trying to solve
this problem and their work is already pushing AI towards higher levels of
intelligence.
turtoni
> turtoni <turt...@fastmail.net> wrote:
> > I think we would be more successful in continuing to work at some
> > "good" questions and perhaps create as a by-product some kind of
> > artificial "intelligence". After all that would appear to be how
> > "intelligence" got started.
>
> > So for example; how do we create an *actual* unlimited supply of
> > renewable energy?
>
> Well, that's a silly question. There's no such thing as "renewable
> energy".
Well, that's a silly answer. There's no such thing as no such thing.
But thats another story.
Anyway: "A natural resource qualifies as a renewable resource if it is
replenished by natural processes at a rate comparable or faster than
its rate of consumption by humans or other users."
> The best energy source around here is the sun, so the real energy problem
> is simply a question of how do we best tap the energy flow from the sun.
> Answering that gives as a few billion years of energy (not renewable).
"The vast power radiated by our sun is generated by the fusion process
wherein light atoms combine with an accompanying release of energy. In
nature, proper conditions for fusion occur only in the interior of
stars. Researchers are attempting to produce the conditions that will
permit fusion to take place on earth."
> > I imagine that from these types of problems a form of artificial
> > intelligence will arise.
>
> Well, the better question along that line is more like:
>
> How do we build a machine that can survive on its own without human help?
How do humans survive without human help? Typically not to well..
> People working on real world robotics questions are already trying to solve
> this problem and their work is already pushing AI towards higher levels of
> intelligence.
Good luck but don't be surprised if other fields of research beat them
to man made artificial intelligence.
> --
> Curt Welch http://CurtWelch.Com/
turtoni
> > I think we would be more successful in continuing to work at some
> > "good" questions and perhaps create as a by-product some kind of
> > artificial "intelligence". After all that would appear to be how
> > "intelligence" got started.
>
> > So for example; how do we create an *actual* unlimited supply of
> > renewable energy?
>
> Well, that's a silly question. There's no such thing as "renewable
> energy".
Well, that's a silly answer. There's no such thing as no such thing.
But thats another story.
Anyway: "A natural resource qualifies as a renewable resource if it is
replenished by natural processes at a rate comparable or faster than
its rate of consumption by humans or other users."
> The best energy source around here is the sun, so the real energy problem
> is simply a question of how do we best tap the energy flow from the sun.
> Answering that gives as a few billion years of energy (not renewable).
"The vast power radiated by our sun is generated by the fusion process
wherein light atoms combine with an accompanying release of energy. In
nature, proper conditions for fusion occur only in the interior of
stars. Researchers are attempting to produce the conditions that will
permit fusion to take place on earth."
> > I imagine that from these types of problems a form of artificial
> > intelligence will arise.
>
> Well, the better question along that line is more like:
>
> How do we build a machine that can survive on its own without human help?
How do humans survive without human help? Typically not to well..
> People working on real world robotics questions are already trying to solve
> this problem and their work is already pushing AI towards higher levels of
> intelligence.
Good luck but don't be surprised if other fields of research beat them
to man made artificial intelligence.
> --
> Curt Welch http://CurtWelch.Com/
Curt Welch
> On Nov 18, 11:48=A0pm, c...@kcwc.com (Curt Welch) wrote:
> > People working on real world robotics questions are already trying to
> > solve
> > this problem and their work is already pushing AI towards higher levels
> > of intelligence.
>
> Good luck but don't be surprised if other fields of research beat them
> to man made artificial intelligence.
Which other fields are you thinking about?
--
Curt Welch http://CurtWelch.Com/
forbi...@msn.com
> I think we would be more successful in continuing to work at some
> "good" questions and perhaps create as a by-product some kind of
> artificial "intelligence". After all that would appear to be how
> "intelligence" got started.
There's certainly a lot of progress to be made. Humans didn't
evolve over night.
> So for example; how do we create an *actual* unlimited supply of
> renewable energy?
or increase the efficiency of our designed systems to the point
where the supply of renewable enery seems unlimited. Disipating
heat can be a problem and humans have a limited opperational
range.
> I imagine that from these types of problems a form of artificial
> intelligence will arise.
Some of us tilt at windmills while others use them to grind
wheat into flour. If I can get a computer to do what I want
easily I'm satisfied.
Alpha
Incorrect; philosophy, in relation to science, is the *question-
generating* or *question-vetting* operation.
> You can't tell which is correct or how correct or incorrect a given
> approach might be.
>
> > Moreover, (reverse) engineering
> > will not solve the problem and may actually lead to many dead ends by
> > just finding ways to go nowhere quicker and better. It will take a new
> > theory and philosophy to do it.
>
> Reverse engineering has already solved it. Many philosophers however don't
> understand this because they have created such a huge cloud of confusion by
> spending so much time debating all the impossible answers they can't get a
> grip on what the truth is.
>
> > Think of it like trying to empirically come up with QED or Relativity w/o
> > any new theory or philosophy of physics.
>
> You have started with the assumption that there is something there
> (consciousness) which is fundamentally hard to understand and explain.
> Your assumption is invalid. Your assumption is created by a simple to
> explain brain function which created in all of us a natural illusion.
Please explain how brain genrates consciousness. Since it is simple
in your mind, it should be simple to tell us how that happens, get it
published in a referred journal (perhaps Journal Of Consciousness
Studies or Brain and Behavior etc.) and thence cvlaim your Nobel.
> If
> you assume the illusion is real, you are left with the hard problem of
> consciousness. If you assume the illusion is only an illusion, then there
> is no problem at all - all hard questions are answered and explained
> leaving a fairly simple material world to understand. By Occam's razor, I
> choose the answer that makes everything simple and answers all the
> questions instead of picking the answer which creates contractions that
> have no answer.
Please answer - oh sage - how brain generates the thought and the
concommitant mind-visual accompanyment: blue cube? I want an exact
specific answer now in terms of how brain's APs or molecules represent
the visualization and the semantic content inherent in : blue cube
thought. Should be easy for you right!
>
> But in philosophy, Occam's razor has no place. All alternatives must be
> explored - as such, you are forced by your very charter to wander endlessly
> into utter silliness. The hardness of the problem attracts you to explore
> the depths of the silliness endlessly hoping to find some "new theory" or
> sudden enlightened understanding which clears it all up.
>
> As an engineer, I have no need to explore such an improbable dead end. If
> I missed something, the philosophers of the world will find it and explain
> it.
You have missed everything if import.
> But after hundreds of years not finding an answer, I'm not holding my
> breath on the expectation that there is something there when all evidence
> suggests there isn't anything there to be found.
But being in the cave, you would have no knowledge of what others
outside the cave have come up with. Or when you do see the shadows of
such on your cave's walls, you dismiss them and skulk back to
reinforcement learning and signals (or wait...is it really all just
"particles" (haha) interacting) as the only thing that exists..
Alpha
OTOH, if we plod along using math tools, without knowing what
questions we want to ask about mind/brain/AI, then the plodding will
continue as it has!
Wolf Kirchmeir
>> All you can do from within philosophy is create multiple possible answers.
>
> Incorrect; philosophy, in relation to science, is the *question-
> generating* or *question-vetting* operation.
>
Yes, that's what philosophers claim, and when it's done right (eg, tests
the logic of the arguments, drilling down to the underlying assumptions,
etc), then it can point up conceptual glitches, which in turn may lead
to useful reformulations the question(s).
But scientists are as good at doing this as are professional
philosophers. Consider the recent attempts to remodel the history of the
universe to eliminate the physically impossible singularity at the
moment of the Big Bang. If space is assumed to be granular (ie, that
space is not a continuum, but that there smallest bits of space), then a
theory that eliminates that singularity is possible. But that assumption
in turn raise the question of what those smallest bits of space are
"embedded in." Which is a metaphor arising from our direct experience of
objects in space.
How can there be smallest bits of something without those smallest bits
being in something else? IOW, how can there be objects without some
space for those objects to inhabit? That's a philosophical question. My
metaphysics allows for it, because I think that "If there is an action,
there must be actor" is fallacious. IOW, "There can be action with no
actor." Analogously, I can accept "There can be objects with no space."
If that sounds muddled, no surprise. It's damn difficult to avoid
thinking in the patterns that language embodies. It's damn difficult to
think in ways that run counter to what language allows as
sensible/sense-making utterances.
"In mathematics, we can know whether what we say is true, but we cannot
know what we are talking about. In poetry, we can know what we are
talking about, but we cannot know whether what we are saying is true."
(after Bertrand Russell.)
"Philosophers are poets who attempt logical proof of their metaphors."
(Wolf Kirchmeir -- you read it here first ;-).)
HTH
--
Wolf Kirchmeir
Alpha
> Alpha wrote:
> > On Nov 17, 5:01 pm, c...@kcwc.com (Curt Welch) wrote:
> [...]
> >> All you can do from within philosophy is create multiple possible answers.
>
> > Incorrect; philosophy, in relation to science, is the *question-
> > generating* or *question-vetting* operation.
>
> [...]
>
> Yes, that's what philosophers claim,
And other scientists; see Bohm's contentions, for example in the
interview responses he gave to Davies in The Ghost in the Atom book.
Many other examples exist through history from thousands of years ago
until present. The first important questions were philosophical;
people theorizing about what actually exists etc. Without empiurivcal
instrrumentalities, that is all that was available in some domains.
Democritus for example. No way to test hios theorizing, but
philosophical speculations abounded based on his ideas. Same occurs
now with QM/QFT; most *thoughtful* physicists want the interpretation
of the theory to tell us, to explain to us ,what relation the
experimental results have to a real world out there. (Of course, some
are willing to just do experiments - but to me that is not very
satisfying.)
> and when it's done right (eg, tests
> the logic of the arguments, drilling down to the underlying assumptions,
> etc), then it can point up conceptual glitches, which in turn may lead
> to useful reformulations the question(s).
Yup.
>
> But scientists are as good at doing this as are professional
> philosophers.
Sure! See the Bohm ref.
>Consider the recent attempts to remodel the history of the
> universe to eliminate the physically impossible singularity at the
> moment of the Big Bang. If space is assumed to be granular (ie, that
> space is not a continuum, but that there smallest bits of space), then a
> theory that eliminates that singularity is possible. But that assumption
> in turn raise the question of what those smallest bits of space are
> "embedded in." Which is a metaphor arising from our direct experience of
> objects in space.
>
> How can there be smallest bits of something without those smallest bits
> being in something else? IOW, how can there be objects without some
> space for those objects to inhabit? That's a philosophical question. My
> metaphysics allows for it, because I think that "If there is an action,
> there must be actor" is fallacious. IOW, "There can be action with no
> actor." Analogously, I can accept "There can be objects with no space."
>
> If that sounds muddled, no surprise. It's damn difficult to avoid
> thinking in the patterns that language embodies. It's damn difficult to
> think in ways that run counter to what language allows as
> sensible/sense-making utterances.
Yes; most transcendental thinkers also have the same difficulties -
they call the things they try to put into language - "ineffable".
>
> "In mathematics, we can know whether what we say is true, but we cannot
> know what we are talking about. In poetry, we can know what we are
> talking about, but we cannot know whether what we are saying is true."
> (after Bertrand Russell.)
>
> "Philosophers are poets who attempt logical proof of their metaphors."
> (Wolf Kirchmeir -- you read it here first ;-).)
;^))
>
> HTH
>
> --
> Wolf Kirchmeir
Josip Almasi
>> property of (sufficiently high) intelligence.
>
> as to require one to immerge from another. Intelligence might just be a
> boundary condition on the scope of consciousness (i.e., awareness). For
> example, a severely retarded person is certainly far less intelligent,
> however, I don't think there is any evidence that they are far less
> conscious. If consciouness emerged (necessarily?) from intelligence then
> shouldn't they be highly correlated?
Fair enough.
Then again, we must not forget social interactions.
Retards get much more attention and training than healthy kids.
Say, I've spent endless hours with a grown up retard (immaturitas
emotionalis, debilitas) unable to remember 3 word sequences. Takes 2
hours for a sequence, 2 hours later he forgets, then all over again;
eventually, he makes it.
There's schools for such people, so society makes up for their disabilities.
Regards...
Isaac
>> > "Isaac" <gro...@sonic.net> wrote in
>> > news:491f9f87$0$33506$742e...@news.sonic.net:
>> >
>> I defy you to contrive a definition of Intelligence that works. For
>> example, using your current definition above, the Earth would be
>> intelligent because it is a system with the capacity to generate
>> solutions (e.g., extremely complex, yet stable atmospheric weather, ocean
>> currents, etc.) to solve novel problems of, for example, maintaining a
>> stable global temperature in the face of many (thousands) changing
>> (novel) variables that are constant obstacles preventing the Earth (Gia?)
>> from attaining her goal of minimizing temperature differences globally.
>
> The earth is intelligent. So is the universe as a whole.
>
> Life looks like it was designed by intelligence because
> it was.
The problem with going this expansive on "intelligence" is that it is not
scientifically useful. For example, with your definition of intelligence a
crystal (esp. while growing) is intelligent. Can you set forth a definition
of intelligence that is scientifically useful and not just philosophically
pleasing?
>The process of evolution is just one more example of the many
> intelligent processes at work in the universe. Evolution is an example of
> a reinforcement learning process
Evolution is not really a pure reinforcement learning process. The fitness
function select viable populations without directly reinforcing any
weightings of the genes. A selection scheme is not a reinforcement scheme.
>and I basically consider all reinforcement
> learning processes to be examples of intelligence.
That is far too broad. Under this scheme a tape recorder, which learns your
voice by reinforcing magnetic monopoles with your voice signal until the
signal to noise ratio reproduces your voice adequately. So, is a tape
recorder intelligent?
>
>> There are many similar examples that use your language but are not
>> considered to be intelligent to anyone reasonable in science. Care to
>> update your definition or defend it?
>
> Many people in science have no clue what they are talking about when they
> use the word "intelligence". As such, they define what is, and what isn't
> intelligent based on total nonsense and ungrounded speculation - as I've
> said before - without using any empirical evidence to argue from.
>
Your solution goes so broad that it is not useful because computers right
now would be intelligent according to your definition, and we know, of
course, they are quite dumb (esp. Microsoft apps :). Please clarify.
> Of course that doesn't stop them, because they like to claim things such
> as
> "subjective experience is outside the scope of empirical evidence".
> And then they tell us what _their_ subjective experience is like and use
> their beliefs about their own subjective experience to "prove" an endless
> list of nonsense ideas about the universe.
>
> The typical argument and thought path starts with the belief that human
> consciousness is something that exists only in humans. Then from there,
> they make the argument that since humans have this magical attribute
> called
> consciousness and other things like the Earth doesn't, that intelligence
> requires consciousness. But since they don't have any clue what creates
> human consciousness, they also don't have any clue what creates
> intelligence and don't really have any way to determine if the earth is
> intelligent or not.
true, but science always has to operate on a best working theory. Yours is
too broad and theirs is too narrow.
>
> And when asked to explain what evidence they have to suggest this
> attribute
> exists only in humans, they use the self serving argument that since they
> "known" it exists in them, and that other humans are physically similar to
> them, that this stuff they known exists in them must also exist in others.
>
> But all that argument and the arguments that grow from it are based on a
> belief that has no support. The belief that "consciousness" is something
> other than simple brain function. That consciousness is not an identity
> with physical brain function.
>
> However, all the empirical evidence we have tells us that assumption is
> wrong. And if we choose to believe what the empirical evidence shows us
> (materialism) - then we know that there is nothing here to explain, other
> than the physical signal processing that happens in the brain which
> produces human behavior.
>
> Once you grasp the significance of what the empirical evidence is telling
> us, all the need of defining intelligence as some sort of link with "being
> conscious" goes away. We are left with defining intelligence is some
> class
> of signal processing algorithm that describes how the brain works. And
> though there are multiple options there, none of them make intelligence
> hard to understand. It's no harder to understand than any typical machine
> learning algorithm for example.
>
> I choose to use the fairly broad and generic definition of intelligence
yes, too broad to be useful for anything but metaphysics.
> being a reinforcement learning system which allows the concept to include
> many processes other than just what the brain does - such as the process
> of
> evolution.
evolution is just an environmental and social selection process of who lives
and dies based on who fits an arbitrary criterion. How is that intelligent?
> You could easily restrict the definition to something closer to what he
> brain does, which would be something more like a real time distributed
> parallel signal processing network trained by reinforcement instead of the
> far broader "all reinforcement learning processes" I like to use.
>
reinforcement learning does not even begin to address the intelligence
problem because it is method of weighting certain nodes in a network more or
less than others, but does not at all address any system level architecture
or algorithm, thus is does not provide a model of reality, just a way to
reinforce a model if you have one. For example, a classic back propagation
neural network (NN) is a implementing reinforcement learning in a NN
architecture with a propagation training algorithm; however, NN have proven
to be completely useless to do anything intelligent and cannot even be made
to converge in a hierarchical configuration even with massive (impractical)
amounts of tagged training data.
<snip>
> Well, I think "goal" is the wrong way to understand the operation of the
> brain though it's not too far off.
>
> The true goal of a reinforcement learning machine is to maximize expected
> future reward. So it's a reward maximizing machine with one prime goal.
>
greedy algorithms are only good for problems with smooth gradient descent;
however, they get trapped at local maxima and minima. Intelligence is
really looking for global maximas so reinforcement learning is a rather dumb
scheme to achieve this. Of course, GA's may try to chaotically hill hop
until a higher hill top is found, but GA's are useless because a suitable
fitness function and gene configuration space are almost impossible to
define; esp., from a top down approach.
> What the prime goal translates into is some internal systems of values for
> all possible behaviors which in turn translates into some behavior
> probability distribution. This in turn must drive whatever mechanism is
> in
> place to select behaviors. The system that decides what behavior to
> select
> for the current context is using the internal system of values to pick
> between alternatives.
>
> Our higher level ideas of "goal seeking" is simply the fall out of a the
> lower level behavior selection system picking the best behaviors for any
> given context.
too vague. See above. A mountain stream searching for the best (i.e., least
energy way) to get to the bottom is implementing all the intelligent
features you propose (e.g., reinforcement of paths that work, "goal
seeking", and selecting locally optimal behaviors), however, saying that a
river defines intelligence is completely useless. Care to be more practical
and specific?
> > When you translate the implementation of the system into a reward
> > trained
> behavior selection system, the frame problem doesn't even make much sense
> to talk about. The frame problem arises nearly as much out of incorrectly
> framing the question of what the purpose of the agent is. However, the
> issues that surround the frame problem are real. But they are all
> answered
> in the context of a system which has the power to prioritize all possible
> responses to stimulus signals. That is, which reaction the system chooses
> at any point in time based on its learned values (priorities if you like)
> is the answer to how the system deals with the frame problem. That is,
> the
> one problem it must solve (how to select which behavior to use at any
> instant in time) is the same answer to the frame problem.
>
> Finding a workable implementation of such a system is the path to solving
> AI.
This is one aspect of the path to AI in the most general sense, but your
ideas are too vague (e.g., you assume an internal model of reality exist to
even be able to perceive what a frame or context is- that does not exist).
> > You said above that goals do not drive perception. That's just not true
> > in
> my view. II think our perception and our behavior selection are one and
> the
> same problem. Perception is a problem of behavior selection.
>
I disagree; e.g., how is perceiving the sound of a drum to be a drum a
"problem of behavior selection"?
> On the sensory input side of the network, the major function is
> perception,
> but as the signals flow through the network, the function transforms into
> behavior selection. So near the output side of the network, it's mostly
> "goal driven" an on the input side it's mostly "perception driven" but I
> believe it's a fairly even continuum though the network as raw sensory
> data
> is translated to raw effector output data.
>
> We see how this works when we test color perception of people raised (aka
> trained) in different cultures with different words for different ranges
> of
> colors. Our perception of color bends to correctly fit the classification
> of light frequency labeled by the words of our language.
>
of course we project our bias on what we perceive; however, that does not
prove that perception is only based on our goal bias, which seems way off
mark to me. Please clarify.
>>
>> >Attention is paid only
>> > to world states which bear on the system's goals (as a background
>> > process).
>>
>> Of course, goals to play an important role in how to focus attention, and
>> to some extent this colors the frame problem, but I do not see how it
>> drives it exclusively as you put it.
>
> It drives it exclusively in my view because behavior selection is all the
> brain is doing and behavior selection works by picking behaviors that are
> estimated to produce maximal expected return for the given context. And
> this general process of selecting the "best" behaviors at the lowest level
> is both the mechanism which creates what we think of as goal seeking and
> the behavior which is think of as attention focus. I see them as one and
> the same process at the low level.
>
the brain models phenomenon far before it forms a goal in relation to that
phenomenon. No doubt they may be organically comingled, but to say your
goal defines your model of phenomenon would amount to a real LSD psychedelic
experience of reality, which the rest of us do not experience. :)
<snip>
> This is because this person has never had experience with this type of
> "frame" in the past and has very little experience with how to correctly
> react to this combination of stimulus signals. None the less, the brain
> will still pick a reaction to the stimulus signal based on the past
> experience the person has had. For this guy, the "reaction" to the frame
> might be to move the eyes to focus on the tree in the background because
> all the city street stuff in the foreground looks mostly like "noise" to
> him.
you are mixing high, system level goals (I want to find some food to eat)
with low level goals (like if you hear an unexpected sound turn your head
towards the sound). It is certainly not useful to say that low level goals
defines our perceptions; of course, they do filter it by pre-selecting what
our higher level systems can become aware of.
<snip>
>> I don't think anyone would say that classic AI would not return to the
>> world to gather more facts to add to its "millions of facts". The issue
>> that Dreyfus says is the problem with AI is that it creates rules that
>> are representations (or symbols) and are compartmentalized, both of which
>> he says the Philosopher Heidegger espouses, which Dreyfus and his set of
>> philosophers/researcher say is not the case. I think every Intelligent
>> system will end up effectively having a constantly evolving set of
>> millions of "rules", so that is not the question. Do you have any
>> counter examples?
>>
>> Cheers!
>> Ariel B.-
>
> I don't fully understand what you are suggesting here because I don't tend
> to read or study the work of the type of people you are studying. I'm not
> sure for example what the debate is on representations.
representations require a modular, hierarchical architecture that parse
realty according to the modules and objectify actions and phenomenons as
being completely separate. Dreyfus and the source he relies upon say the
brain is a flat network with no modules and have no separation between
action and phenomenons.
>
<snip>
>
> However, just because the GOFAI approach ran into a wall after awhile,
> that
> doesn't mean that using "symbols" to represent something was wrong.
>
Dreyfus says using "symbols" in any way is very wrong.
> The brain uses a common language of symbols to represent everything as
> well. Those symbols however are spikes. Digital computers use 1 and 0
> symbols to represent things. Manipulating symbols which are
> representations is exactly what the brain is doing. Any argument to the
> contrary is misguided.
>
OK, explain, then, how neural networks (NN) have explicitly symbols as you
say is mandatory in an intelligent system.
> The solution to creating human like behavior in a machine is to build
> symbol manipulating machines (aka signal processors), but the symbols must
> be a level closer to spikes or bits, than to English words.
>
> I also think that the correct implementation is along the line of a
> confectionist network which is processing multiple parallel signal flows.
> So from that perspective, I think it's more useful to think of the network
> as a signal processing machine than as "representations with symbols".
OK, then address the NN question above. Generally, I believe, you need a
hierarchy to represent a symbol. Dreyfus et. al. says the NN is flat.
>But
> it's the same thing no matter which way you talk about it. An AM radio
> signal is still a representation of the vibration of the air, and is also
> a
> representation of the thoughts of the DJ which was speaking on the radio.
Not a good example, since the AM radio has no symbols to represent the voice
info. It is just a amplitude modulated analog signal.
> How you choose to label these systems is matter of viewpoint far more than
> a true matter of what the system is doing or how it works.
Not really, see my above comments on these systems you outline don't work as
you stipulate.
Thanks for your thoughtful reply!
Cheers!
Ariel-
Curt Welch
Sure it does. You just have to learn to look at it correctly.
If you look at the entire population of a single species as a large
distributed reinforcement learning machine (instead of being focused on a
single individual), then you can see that the weighting of the genes
happens at the gene pool level. In the population, at any moment in time,
the current population of genes and alleles is the waiting of that gene in
the population.
Each birth, and death, changes the weighting of the genes in the gene pool.
It's very much a reinforcement learning machine.
> >and I basically consider all reinforcement
> > learning processes to be examples of intelligence.
>
> That is far too broad. Under this scheme a tape recorder, which learns
> your voice by reinforcing magnetic monopoles with your voice signal until
> the signal to noise ratio reproduces your voice adequately. So, is a
> tape recorder intelligent?
I get made fun here all the time of for saying rocks are conscious. :)
Yes, you can consider that aspect of the the tape recorder an example of
intelligence as well. But it's a stretch. There's not a good example of
the machine learning slowly by multiple reinforcement events. You don't
have to train the recorder by recording the same voice 10 times in order to
train it to playback your recording correctly. It always happens in one
step - which means it learns so fast that it has almost no memory of the
past - which makes it a very low level of intelligence.
Yes, the idea of reinforcement learning being intelligence is very broad,
but I think it fits. I think it is the process people are in fact talking
about when they talk about intelligence even though most don't understand
that.
The problem here is that most people have no clue what type of process is
at work creating human behavior and they just consider it "undefined magic"
which they use the name "intelligence" to label. And since they don't see
human like behavior coming from anything else in the universe, and no one
has been able to make human like behavior come out of our machines despite
no end of effort to do so by lots of very smart people, they make the
obvious assumption that the process behind human behavior must be something
very complex and hard to understand. As such, something so simple as
reinforcement learning seems to be an obvious wrong answer. But it's not.
This confusion over consciousness also makes the answer hard to accept.
> >> There are many similar examples that use your language but are not
> >> considered to be intelligent to anyone reasonable in science. Care to
> >> update your definition or defend it?
> >
> > Many people in science have no clue what they are talking about when
> > they use the word "intelligence". As such, they define what is, and
> > what isn't intelligent based on total nonsense and ungrounded
> > speculation - as I've said before - without using any empirical
> > evidence to argue from.
> >
>
> Your solution goes so broad that it is not useful because computers right
> now would be intelligent according to your definition,
They are.
> and we know, of
> course, they are quite dumb (esp. Microsoft apps :). Please clarify.
TD-Gammon (an example I use all the time here) is a reinforcement trained
program that plays the game of backgammon. It leaned, on it's own, now to
best play the game, by playing itself in millions of games. It plays at or
near the level of the best human players, and it has in one example, even
taught the humans how to better play a certain type of opening move in the
game.
This program is interesting for many reasons - but the most notable is
simply that it leaned how to win backgammon games on it's own, instead of
having the author of the program hand-code all his knowledge about how to
play the game into the program. He built a learning machine, and let it
learn on it's own how to play. The same guy had written other backgammon
programs - I believe his programs were the strongest backgammon programs in
the past. But with TD-Gammon, he took a different approach and removed all
the typical strategy and rules he built into his past programs, and instead
just built a pure learning machine with no initial knowledge about how to
pick the best move. It learned how to play the game using a reinforcement
learning algorithm which was adjusting the weights of neural network which
act as it's behavior selection system. A neural network that only has a
few hundred weights total, was able to hold all the knowledge needed to
play backgammon at the level of the strongest human players.
Most the machines we build don't learn on there own. There is very little
learning happening in Microsoft windows for example. Though we do see it
coming into our applications more and more over time, like with spam
filters, and spelling correction.
The bottom line here is that most humans think humans are something very
special and somewhat magical. Many people even think that duplicating
human behavior in a machine is impossible.
I hold the view that it's actually far far simpler than most believe. I
think our intelligent behavior is nothing more than what emerges from a
fairly simple reinforcement trained neural network which translates sensory
inputs effectors inputs using real time temporal processing techniques.
Unlike most, I don't believe the underlying technology of the neocortex is
a lot of different custom built modules each with a different purpose all
designed and wired by millions of years of evolution. I think it's in
general on homogeneous type of learning network which is simply used to
process different data in different locations - like our memory chips in
our computers are all identical, but each have a different function based
only what is stored in them.
TD-gammon is a working example of how intelligent behavior can be produced
by the application of reinforcement training on a neural network. It's
network only has something like 40 nodes, but yet that's enough to store
all the knowledge needed to play backgammon better than 99% of all humans.
What do you think could happens when you figure out how to do something
similar with 100 billion nodes like the brain is using instead of 40?
The network implementation used by TD-Gammon isn't a real time reaction
machine like the brain is. It's a program that picks moves in a discrete
game. It can't be scaled up to solve human behavior problems. But it shows
how close we already are.
And yes, I already consider programs like TD-Gammon to be good examples of
true machine intelligence.
> > Of course that doesn't stop them, because they like to claim things
> > such as
> > "subjective experience is outside the scope of empirical evidence".
> > And then they tell us what _their_ subjective experience is like and
> > use their beliefs about their own subjective experience to "prove" an
> > endless list of nonsense ideas about the universe.
> >
> > The typical argument and thought path starts with the belief that human
> > consciousness is something that exists only in humans. Then from
> > there, they make the argument that since humans have this magical
> > attribute called
> > consciousness and other things like the Earth doesn't, that
> > intelligence requires consciousness. But since they don't have any
> > clue what creates human consciousness, they also don't have any clue
> > what creates intelligence and don't really have any way to determine if
> > the earth is intelligent or not.
>
> true, but science always has to operate on a best working theory. Yours
> is too broad and theirs is too narrow.
I don't just stop at "reinforcement learning".
To duplicate human behavior in a machine, I believe we need a real time
confectionist network that can process parallel data streams (aka like the
brain, but also like our ANNs) which is trained by reinforcement.
I have simple working examples of multilevel networks that can be trained
by reinforcement - thought they still lack important learning powers.
I've talked about them in detail here in c.a.p. in the past many times over
the years.
I believe networks something like these are the key to creating AI.
Because it's another example of the fundamental process of reinforcement
learning at work. Most people won't understand this connection until
someone builds a working AI machine that actually acts somewhat human-like,
and then explains to them how that machine works.
> > You could easily restrict the definition to something closer to what he
> > brain does, which would be something more like a real time distributed
> > parallel signal processing network trained by reinforcement instead of
> > the far broader "all reinforcement learning processes" I like to use.
> >
> reinforcement learning does not even begin to address the intelligence
> problem because it is method of weighting certain nodes in a network more
> or less than others, but does not at all address any system level
> architecture or algorithm, thus is does not provide a model of reality,
> just a way to reinforce a model if you have one. For example, a classic
> back propagation neural network (NN) is a implementing reinforcement
> learning in a NN architecture with a propagation training algorithm;
> however, NN have proven to be completely useless to do anything
> intelligent and cannot even be made to converge in a hierarchical
> configuration even with massive (impractical) amounts of tagged training
> data.
Your error is in believing their uselessness has been proved. No such
proof exists. What you are pointing to is the fact that finding the right
architecture is hard.
Back prob is NOT reinforcement learning. Back prop is learning by example,
or sometimes called supervised learning which puts a severe limit on
creativity that does not exist with true reinforcement learning.
Reinforcement learning is trial and error learning. It's learning by
experimentation. Back prob is learning by example where the examples are
normally provided by some external teacher who must always be "smarter"
than the "student". Reinforcement learning allows the student to learn
things the teacher never knew (as per what happaned with TD-Gammon).
> <snip>
>
> > Well, I think "goal" is the wrong way to understand the operation of
> > the brain though it's not too far off.
> >
> > The true goal of a reinforcement learning machine is to maximize
> > expected future reward. So it's a reward maximizing machine with one
> > prime goal.
> >
>
> greedy algorithms are only good for problems with smooth gradient
> descent; however, they get trapped at local maxima and minima.
> Intelligence is really looking for global maximas so reinforcement
> learning is a rather dumb scheme to achieve this. Of course, GA's may
> try to chaotically hill hop until a higher hill top is found, but GA's
> are useless because a suitable fitness function and gene configuration
> space are almost impossible to define; esp., from a top down approach.
Yes, the problem of not getting trapped at a local maxima is important.
But I believe the solution to that is to fracture the problem into a huge
number of smaller problems. That is, each node in a network is solving its
own little piece of the problem. And though many nodes at any one time
might be trapped at a local maxima, as long as some nodes are not trapped,
they can still learn and improve the networks solution. And as those
changes, it ends up changing the landscape for other nodes, freeing some of
them to make progress. As long as some nodes are still making progress,
the network as a whole can still make progress.
Another way to look at this is that the learning is not just a one
dimensional or two dimensional hill climbing problem. If you have 100,000
nodes all learning in parallel, it's a 100,000 dimension hill climbing
problem. The more dimensions you can break the problem into, the more
likely they will be a path around all the local maxima traps.
The key to making this work, is finding the right architecture. The
architecture you apply to the problem determines the nature of the
multidimensional surface the "hill climbing" is applied to. The
architecture you pick defines the nature of the problem it's trying to
solve.
> > What the prime goal translates into is some internal systems of values
> > for all possible behaviors which in turn translates into some behavior
> > probability distribution. This in turn must drive whatever mechanism
> > is in
> > place to select behaviors. The system that decides what behavior to
> > select
> > for the current context is using the internal system of values to pick
> > between alternatives.
> >
> > Our higher level ideas of "goal seeking" is simply the fall out of a
> > the lower level behavior selection system picking the best behaviors
> > for any given context.
>
> too vague. See above. A mountain stream searching for the best (i.e.,
> least energy way) to get to the bottom is implementing all the
> intelligent features you propose (e.g., reinforcement of paths that work,
> "goal seeking", and selecting locally optimal behaviors), however, saying
> that a river defines intelligence is completely useless. Care to be more
> practical and specific?
Yeah, but this message is too long already. Let me post another one to
talk about my approach in more detail.
Again, let me explain my approach from the bottom up in a different message
and I think you will better understand what I'm getting at there.
:)
In my networks, there are two processes at work which are commingled (to
use your term). One process is learning the statistical properties of the
data (unsupervised learning from sensory data), and the other is the
reinforcement learning process which is applied on top of the statistical
system. The idea of the statistical system is to compress the data to
allow the internal signals of the network to represent as much information
about the state of the environment as possible with as little redundancy as
possible. It must also create a mapping that solves the invariant
representations problem. That creates the default mapping of the network.
Reinforcement learning is then added on top of that to warp the default
mapping in order to maximize reward.
To say the goal "defines" the model, is simply to say that the
reinforcement training re-defines the default model of the system based on
experience.
> <snip>
>
> > This is because this person has never had experience with this type of
> > "frame" in the past and has very little experience with how to
> > correctly react to this combination of stimulus signals. None the
> > less, the brain will still pick a reaction to the stimulus signal based
> > on the past experience the person has had. For this guy, the
> > "reaction" to the frame might be to move the eyes to focus on the tree
> > in the background because all the city street stuff in the foreground
> > looks mostly like "noise" to him.
>
> you are mixing high, system level goals (I want to find some food to eat)
> with low level goals (like if you hear an unexpected sound turn your head
> towards the sound). It is certainly not useful to say that low level
> goals defines our perceptions; of course, they do filter it by
> pre-selecting what our higher level systems can become aware of.
Though we like to talk about these effects using different concepts at the
high levels and low levels, I believe they are mostly just the same effect
at work at different levels. I believe all the high level emerge "goals"
we talk about in human behavior are simply emergent side effects of the low
level system seeking to maximize the expected future reward signal.
> <snip>
>
> >> I don't think anyone would say that classic AI would not return to the
> >> world to gather more facts to add to its "millions of facts". The
> >> issue that Dreyfus says is the problem with AI is that it creates
> >> rules that are representations (or symbols) and are compartmentalized,
> >> both of which he says the Philosopher Heidegger espouses, which
> >> Dreyfus and his set of philosophers/researcher say is not the case. I
> >> think every Intelligent system will end up effectively having a
> >> constantly evolving set of millions of "rules", so that is not the
> >> question. Do you have any counter examples?
> >>
> >> Cheers!
> >> Ariel B.-
> >
> > I don't fully understand what you are suggesting here because I don't
> > tend to read or study the work of the type of people you are studying.
> > I'm not sure for example what the debate is on representations.
>
> representations require a modular, hierarchical architecture that parse
> realty according to the modules and objectify actions and phenomenons as
> being completely separate. Dreyfus and the source he relies upon say the
> brain is a flat network with no modules and have no separation between
> action and phenomenons.
I suspect I agree with the general direction Dreyfus is headed (and
Freeman), but I happen to believe that even in the type of network he's
thinking about, sections of that flat network will naturally tend to form
what we would likely call modules and a hierarchical architecture.
> <snip>
>
> >
> > However, just because the GOFAI approach ran into a wall after awhile,
> > that
> > doesn't mean that using "symbols" to represent something was wrong.
> >
> Dreyfus says using "symbols" in any way is very wrong.
Yeah, I've debated the fact that pulses really are symbols. There are some
here that call me an idiot for suggesting that. (but lots of people like to
call me an idiot).
> > The brain uses a common language of symbols to represent everything as
> > well. Those symbols however are spikes. Digital computers use 1 and 0
> > symbols to represent things. Manipulating symbols which are
> > representations is exactly what the brain is doing. Any argument to
> > the contrary is misguided.
>
> OK, explain, then, how neural networks (NN) have explicitly symbols as
> you say is mandatory in an intelligent system.
Spikes are symbols. If you have a light sensor which generates a frequency
modulated spike train based on light intensity, you can think of each spike
the sensor generates as being equivalent to a person yelling a word which
means "I just saw a billion more photons!". The pulse is clearly a
symbolic representation of a physical event detected by the sensor.
This clearly fits in my mind the definition of a symbol:
http://en.wikipedia.org/wiki/Symbol
A symbol is something such as an object, picture, written word, a sound,
or particular mark — that represents something else by association,
resemblance, or convention,
Likewise, if these symbols are processed by neurons an produce new symbols
based on some logic as a function of other symbols, the new symbol (spike)
takes on new meaning because the condition the which causes the generation
of the symbol will be different - it will represent something else.
It's possible that intelilgence could be created with an analog signal
system in which case we wouldn't really be able to use the word "symbol"
anymore because symbol implies a discrete representation system. Though I
suspect all our best attempts to create a learning system that duplicates
human-like learning and behavior skills will be based on discrete symbols.
I say this because I suspect it's very hard to do high quality
reinforcement learning if the lowest level representation system isn't
using discrete symbols. That is, it seems to me there must be discrete
events that are being conditioned. I certainly have only played with
discrete signaling systems though I would not be surprised if one day
someone figured out how to do it in a purely analog domain.
> > The solution to creating human like behavior in a machine is to build
> > symbol manipulating machines (aka signal processors), but the symbols
> > must be a level closer to spikes or bits, than to English words.
> >
> > I also think that the correct implementation is along the line of a
> > confectionist network which is processing multiple parallel signal
> > flows. So from that perspective, I think it's more useful to think of
> > the network as a signal processing machine than as "representations
> > with symbols".
>
> OK, then address the NN question above. Generally, I believe, you need a
> hierarchy to represent a symbol. Dreyfus et. al. says the NN is flat.
I'm going to address this in a follow up....
> >But
> > it's the same thing no matter which way you talk about it. An AM radio
> > signal is still a representation of the vibration of the air, and is
> > also a
> > representation of the thoughts of the DJ which was speaking on the
> > radio.
>
> Not a good example, since the AM radio has no symbols to represent the
> voice info. It is just a amplitude modulated analog signal.
Right it's an analog representation instead of a discrete representation.
It very much is a representation, but we don't don't use the word "symbol"
unless it's a discrete representation system.
This fact that the English word "symbol" only applies to discrete
representation systems says far more about English that it says about how
the machines work.
> > How you choose to label these systems is matter of viewpoint far more
> > than a true matter of what the system is doing or how it works.
>
> Not really, see my above comments on these systems you outline don't work
> as you stipulate.
>
> Thanks for your thoughtful reply!
>
> Cheers!
> Ariel-
I'll follow up and talk a bit more about the specifics of the type of
networks I've been playing with to try and solve the problem of general AI
by building a reinforcement trained learning network.
Curt Welch
> "Isaac" <gro...@sonic.net> wrote:
> > "Curt Welch" <cu...@kcwc.com> wrote in message
> > news:20081119002253.348$D...@newsreader.com...
> > I disagree; e.g., how is perceiving the sound of a drum to be a drum a
> > "problem of behavior selection"?
>
> Again, let me explain my approach from the bottom up in a different
> message and I think you will better understand what I'm getting at there.
Ok, so let me give a bit more of an overview of the type of networks I've
been playing with and how I think this approach is going to solve AI.
First, let me say I just play with this a hobby. But it's a hobby I've
been messing with for about 35 years now (on and off - mostly off - I can
ignore it for years and then come back to it and work hard at new idea for
weeks on end).
For years I played mostly with feed forward binary networks where every
node in the multilayer network would calculate a new output value for each
computation cycle of the network. My goal has always been to understand
how to apply reinforcement learning to a multilayer neural networks. IT's
hard to get these sorts of network to do anything interesting - but I've
had some success over the years and far more failures.
But a few years back, I got interested in the idea of simulating async
pulse networks instead of the binary networks. That was quite a paradigm
shift because all the data existed in the temporal domain. No longer would
a node be given a set of "input values" to computer a value from. Pulses
would show up at the inputs at different times (like a real brain) with
synchronization of the pulses. This shifted my view on the type of
function the nodes of the network would perform. They would need memory of
how much time had passed since previous pulses and their behavior would be
a function of those sorts of time values. It was a whole new way of
thinking about networks for me.
In addition, there was a recurring problem I had in the old style networks
where there was a constant need for conservation of information in the
network. This was a need driven by the fact I was trying to train these
networks using only reinforcement - which means the network as a whole was
given rewards, but no indication as to what it was that the network did
which cased the reward.
If I was trying to train the network to perform some simple logic function
on the inputs, there was this complexity that if the input data that was
needed to generate the correct output was "lost" in the lower levels of the
network, and never made it to the final output levels, any reward the
network got for doing what looked like the right thing, would never in fact
be the true correct behavior because even though the output happened to be
correct, it was always correct for the wrong reason.
None the less, this led me to realize that for these types of systems, it
was important to try and make sure all outputs were a function of all
inputs and that the network didn't create functions which in effect dropped
input values from the function (as if they were multiplied by a value of 0
for example).
This sort of information conservation need, combined with the idea of using
async pulse signals, got me to the idea of using a pulse sorting network.
That is, in stead of thinking of the problem as network nodes that "fired"
based on different logic, I started to look at it as a problem of routing a
single pulse though some path in the network, and then reaching an output.
The pulse was not allowed to fork, or to die out. It was forced by the
design of the system to always be routed to some path.
With this idea, I've been using playing with networks that have nodes that
have a single input, and two outputs. They act as binary switches. Each
time they receive a pulse, they must make a binary decision about which way
to route the pulse. The networks I use are feed forward, so there's no
direct loops in them. The pulse enters, flows through the network, and
exists somewhere.
In software, I implement these networks so that they only route one pulse
at a time. It makes the performance of this type of network greatly
outperform any typical NN I've used in the past which required that every
node in the network have it's output value recalculated for every cycle of
the network. In this new approach, only the nodes that pulse passes though
have to be dealt with.
Now, even though the network passes pulses in a feed-forward fashion, there
can be feedback loops in multiple ways. First, the logic a node uses to
determine how to route a pulse is based on pulses that have passed though
the network in the recent past. They can look at how much time has passed
since a node flowed though a down-stream node in the network for example.
Doing this, creates a feedback loop from the higher (downstream) levels of
the network to previous levels. So even though the pulses don't travel in
the feedback loop, the information about the pulse effectively does,
creating feedback effects.
How each pulse gets routed through the network, then becomes a complex
function of both how the network is wired and configured, and on which
pulses passed through which parts of the rest of the network, in the recent
past.
One way to look at this type network is as a big decision tree - except
it's a mesh instead of a tree and it has multiple inputs and multiple
outputs. For each pulse injected into the network, the individual nodes
all make a decision about how to route the pulse, and the combined effect
of all the nodes working decision, creates a network-wide decision on how
to map a single input pulse to a single output path.
This type of architecture lends itself very nicely to the domain of
reinforcement learning. That's because the nodes are producing a clear
binary behavior. If the behavior of routing a pulse to one output is
punished, it's clear what the alternative is - increase the probability of
routing it the other way in the future. If the behavior has multiple
options, reinforcement learning requires that the a statical value must be
traced for each option. With binary decisions, a single statistic can be
maintained for the single binary decision (with the statistic representing
the probability of one decision, and the probability of the alternate
decision just being 1-p.
Now the idea of this type network is that is that for each pulse received,
the network must make an instant decision about how to react to it - about
which output to send the pulse to. The decision is not however made by the
entire network. It's only made by the nodes the pulse passes though. So
how the pulse is routed thought the network is really creating two effects
at once. Not only is each node that routes the pulse to some downstream
node making part of the decision, it's doing it by deciding which
downstream node (or downstream sub-network logically) should make the rest
of the decision. Each node then is expected to learn which downstream sub
network, works best at producing higher rewards, based on the current
context (as best as as single node "understands" the current context". So
the learning problem gets distributed over the network, with different
parts of the network specializing in different contexts (in theory).
Now, the idea I'm trying to create, is that the routing logic of the
network needs to, by default, attempt to solve the invariant representation
problem. That is, by using temporal correlations in the data, it should
build a default parsing network, that tends to assign different nodes to
represent different invariant contexts of the sensory data.
To be a little simpler, this means that such a network should, if it does
what I think it needs to do, learn to recognize common temporal patterns in
the data, by routing the pulses that represent that pattern, to common
nodes. That is, if it learns to recognize dogs, this means the network
will learn to route the pulses which represent a dog, to the "dog" node
someone down in the higher levels of the network.
This is not something it will be trained by reinforcement to do. It is
something it must be able to do simply by studying the temporal predictive
nature of the raw signals alone. It's information that exists in the input
data, and that information is what guide the network into correct "parsing"
sensory data into invariant representations. Those invariant
representations are the indication of what state the environment is
currently in.
So, without any reinforcement learning, the network must be able to
configure itself to correctly model the state of the environment as
represented in the sensory data. This alone should cause the final outputs
of the network to represent high level features of the environment. I've
not yet found the correct logic to make this work like I needs to by the
way - and it's the part I generally spend my time thinking about.
With that much working, the network in effect has learned to "see" the
invariant objects that exist in the environment. The larger the network,
the higher resolution the system has - aka the more "things" it has the
power to "see" in the environment. A key concept here is that the outputs
are assigned so they are equally likely in probability. So the feature
assignment works so as to make each feature the network learns to recognize
equally likely to show up in the systems environment.
Now, on top of this we add the all important reinforcement learning.
Without reinforcement learning at work, we can build a network that learns
to recognize a dog, but it has no clue, how it should react to a dog. It
has no way to determine good from bad - it would have no values. Should it
kill the dog, eat it, run away from it? Without reinforcement learning, it
might have the power to predict how the dog will react to any action
created by the agent, but it would have no way to prioritize one action
over another.
So the unsupervised learning gives it the power to "understand" the world,
but the reinforcement learning tells it how to react to the world. In
short, it picks the behaviors which are expected to produce the highest
levels of future rewards.
Now, as I pointed out above, a key part of the unsupervised learning is
that it naturally divides the environment into equal probability features.
Each signal path in the network represents a different feature of the
environment but each feature tends to show up over time with roughly equal
probability. The output features - the one we see at the output signals of
the network, are in fact the behavior of the system.
So, if a given output from the network was wired to a cause the robot to
lift it's right arm, and by chance, that "feature" happened to be the "dog"
feature, this robot by default would lift it's left arm whenever it saw a
dog. That would be an odd thing do to, but that might be the default
behavior of this network.
We add reinforcement learning on top of this, to select how the network
should really respond to dogs.
To do this, we simply bend the probability of each feature. Recall that
each node in the network had two outputs. This also means each node in
effect is classifying each pulse it receives, as being on of two features.
Each node in effect is acting as a feature extractor.
Let me give a real example to make this clearer. If a node was fed a pulse
stream from a light detector, it could receive high frequency pulses for
bright light, and a stream of low frequency pulses for dim light. The
frequency would measure the current intensity of the light - or, the
temporal spacing of the pulses would indicate the light level. A node
which remembers how much time has passed since the last pulse, could use
that to make it's classification decision (and this is a network design
I've played a lot with). It would sort the pulse out one path if the
previous pulse was close in time (high frequency) and sort it out the other
path if the previous pulse was further in the past (low frequency). The
pulses coming into the node in effect have a symbolic meaning of "light".
But the pulse going out the node, have gained addition symbolic meaning.
Node out one side would in effect mean "bright light", and nodes out the
other side would in effect mean "dim light".
So we see with this example how a node can classify pulses by sorting them.
One output signal is an indication of "low light level" and the other is an
indication of "high light level".
If we then route the low light pulses to an output that makes a simple two
wheel robot turn right, the network will cause the robot to turn right when
it sensed a low light condition. If the network also routed the high light
level pulses to the "turn left" output, the same robot would turn left in
response to bright light.
Now, when classifying light as "bright" and "dim" the default behavior of
the node is to set the dividing line so that the "bight" output is active
on average as much as the "dim" output is active (aka same number of pulses
over time on average).
To add reinforcement learning on top of this, we allow the rewards received
by the network, to shift that default balance. If the network gets more
rewards over time for sending pulses out the "bight light" side, we shift
the classification behavior of the node to send more pulses out that side.
It does this by lowering it's definition of how much light is needed to be
considered "bright" (or, we think of it as what pulse frequency the node
uses as the dividing line between it's definition of dim and bright light).
This ability to shift where the pulse are being sent, allows such a network
to route a signal created by the network, to any output. If it needs to,
it just shifts it's probability so as to route all pulses out one side, and
none out the other. It would do that if one output consistently produced
more rewards than the other. A small network like this can, for example
produce these "bight light", and "dim light" signals in the first layer,
and then using a few more layers, route one signal to the "turn left"
output, and the other signal to the "turn right" output.
A network like this, combined with two light sensors (eyes) - one on each
side of the head of the robot - could allow such a network to make the
robot turn towards the brighter light. That is, it can learn to be a
simple light seeking robot by how the pulses are classified, and which
outputs they get sent to.
When we look at the outputs of such a network, we will find that the signal
being sent to the "turn right output" can be thought of as "brighter light
on right side of robot". So looking back into the network at what what
caused the signal to be created, we can say the signal "means" "bight right
on right side of robot". But looking forward at what the signal will do,
we can say the "meaning" of the pulse is a command the the robot to "turn
right" (aka spin left wheel faster ans right wheel slower).
Such a network can be said to be "perceiving that the light is brighter on
the right side of the robot". But it learned to sense this condition,
based on the reinforcement learning that created, and tuned this signal, to
have that meaning, because that's the signal which allowed the robot to get
the most rewards from the environment.
In this network design, every node in the network is performing the dual
role of feature detections, and behavior selection.
In this type of network, "perception" and "behavior selection" are one and
the same thing.
In fact, if you look at this example, we can say that the network output
which makes the robot turn right, is the networks _perception_ of when it's
time to turn right. Or, the networks perception, of when it thinks it
needs to turn right.
In fact, the purpose of the entire network, is to transform the raw sensory
data signals, into the correct "perception" of when it should act. All the
intermediate signals in the middle layers of the network, end up being
there only because they help this multilayer network compute the correct
output values.
By looking at what happens as the sensory data flows forward thought the
network, we can all the process an act of perception. But by looking at it
from the outputs backward, we can call the same function a process of
"action production".
I actually spend a few years playing with network designs that performed
the perception task in one network, and the action selection (reinforcement
learning) task in a second module. I eventually figure out that can't
work. Trying to do it separately creates a scaling problem that's
unworkable. Or, maybe better said, doing both in each node at the same
time, improves the power of the multilayer network to scale exponentially.
I strongly suspect this same effect happens in the human brain because of
the scaling requirement. Though we are taught to think of percpetion as
being different than behavior selection, I strong suspect, that in the
brain, like in my networks, it's actually the same thing.
> > the brain models phenomenon far before it forms a goal in relation to
> > that phenomenon. No doubt they may be organically comingled, but to
> > say your goal defines your model of phenomenon would amount to a real
> > LSD psychedelic experience of reality, which the rest of us do not
> > experience. :)
Right. A rat can't learn to respond to the stimulus of a flashing light
unless it can first "see" the flashing light! From my description above,
you should now understand that the approach I'm looking at must include an
unsupervised learning technique that allows the network to configure itself
so that it produces internal signals that represent things like that light,
before reinforcement learning is used to figure out what to "do" with the
signal.
SO I agree completely with what you are saying there. And my approach is
as you say to "commingle" the unsupervised learning which creates the
foundation of the networks ability to correctly perceive the environmental
(correctly parse it into invariant features), with reinforcement learning
to "bend" that parsing by making classification categories larger and
smaller, to create optimal behaviors (behaviors that produce higher
rewards).
> > you are mixing high, system level goals (I want to find some food to
> > eat) with low level goals (like if you hear an unexpected sound turn
> > your head towards the sound). It is certainly not useful to say that
> > low level goals defines our perceptions;
Well, based on what I wrote above, you might grasp that I think it's not
only useful to say that, but it's required.
> > of course, they do filter it
> > by pre-selecting what our higher level systems can become aware of.
Right. You might think of it as "focus", or "attention". I just see it as
"behavior selection". Do you grasp how that works now? If the perception
network and behavior selection network are one and the same, then as the
system learns the correct behavior, it's also learning to bend it's
perception at the same time, and it's also learning to "filter out" what it
"wants" while "throwing away" what it doesn't want.
Though we have lots of different ways to talk about these things at the
high level, I think the type of networks I'm playing with, does it all at
the same time as one unified approach to the problem.
> > OK, then address the NN question above. Generally, I believe, you need
> > a hierarchy to represent a symbol. Dreyfus et. al. says the NN is
> > flat.
Ok, so in my network, I use a multilevel network that creates a hierarchy
to parse and transform sensory data, into effector outputs. I only talked
about how signals were split apart by the switching function, but didn't
touch on how they are merged together. Pulses can merge in this type
network simply by wiring two outputs to a single node input. But how and
why they should merge is an issue I still struggle with. We will ignore
that fact I don't have that correct logic figured out yet....
The result however is that the network, like any neural network, useses a
hiearchy of signals to create higher level more complex siganls the deeper
you go into the network. In theory, the network might have an input that
means "light detected at pixel location 100,345 in the right eye", but as
the pulse travels though this classification network, it accumulates
additional "meaning". Like in my example above twhere it transformed from
"light" at one level to "dim light" at the next level. The idea is that
after multiple levels it might take on the meaning (based on the node it
reaches) as meaning "dim light which is part of sharp edge next to 30 deg
corner .... which is part of an animal ear which is on the right side of an
animal head which is part of the dog...".
So throughout the network each output of each node in the network
represents some feature of the input signals - aka some features of the
environment. And the deeper you go into the network, the higher you are
going in a hierarchy which defines large and more complex features -
starting at "light" at the lowest level, and hitting "dog" at a much higher
level in this example.
So that's the basic hierarchy the low level network is using to define it's
symbols with this type of network (and again, I think the brain is doing
some thing basically the same).
I'm using a "flat" network to do this, but it's a multilayer "flat" network
with lots of freedom to make cross connections and use feedback from higher
levels to lower levels to correctly "parse" the sensory data. I don't know
for what Dreyfus means by "flat".
> > > How you choose to label these systems is matter of viewpoint far more
> > > than a true matter of what the system is doing or how it works.
> >
> > Not really, see my above comments on these systems you outline don't
> > work as you stipulate.
:)
My current networks only have limited success. The thing my current
network is not doing correctly is the hardest part of this direction I'm
trying to go which is to get the unsupervised learning working correctly.
I only abstractly understand what I think it needs to do, and I have not
yet translated those abstract ideas into working code. But it seems like
the right approach to me so I continue to tackle it from this direction.
And the overall architecture of using pulse sorting as the underlying
paradigm still looks very good to me, so I'm sticking with that approach
for now to see what I can make work.
There's another important aspect to this design I didn't cover above.
This type network has the power to learn how to correctly respond to
stimulus inputs, and it's response is not just a function of what is
happening at one instant in time, but what has happened over the past few
minutes in the network - that is, which nodes have had pulses pass thought
them. So, for a high level example, if there's a "dog" node in the network,
how the network will route a current pulse, can depend on whether the
network has routed a lot of pulses though the dog node in the recent past
(past few seconds or minutes). So this means, it can perform a function
such as "when we see X, if there's a dog around in the last 10 seconds, do
Y, otherwise, do Z".
So though the network is structured to always answer the immediate question
of "what is the best way to respond to the input I just received", it's
answer to how to best respond is always a function of what has recently
been seen.
Such a system can learn a large set of reactions to different environmental
conditions, and the effect is that it "strings together" lots of little
reactions, to create long complex chains of reactions. Each pulse it
"sorts" is the "little reaction" in this case. And of course, if it's got
some high volume sensory data to deal with, it might be sorting thousands
or even millions of pulses every second to "string together" a long complex
string of behaviors. But to produce a complex behavior sequence, such a
system must be driven by a constant flow of "clues" from the environment.
What this architecture (as I've explained so far) doesn't explain, is how
the system could develop goal based behavior. So far, the system is
basically forced to react to the same environment the same way every time.
Every time it sees a specific blue chair, it performs the "sit down"
behavior for example because so far, as I've described the architecture,
it's forced to respond to same environment the same way every time.
The solution to this, is to add global feedback from the networks outputs,
back to another set of inputs. What this does, is allow the network to
"sense" what it's been doing. So if there's an output that makes the right
wheel turn faster, that signal is sent back into the network as yet another
separate sensory input. The network processes these inputs just like it
processes all inputs - it applies the unsupervised learning rules to the
data to create a feature extraction network. This allows it to "see"
features of it's own behavior, and to react to what it's been doing in the
past.
This feedback path allows the system to react only to its own behavior, and
ignore the environment if it needs to. Or better said, this feedback path
allows the system to see it's own behavior as part of the state of the
environment that it is reacting to.
With this feedback, the system can learn a set of reactions to do something
like this: Turn right for 10 seconds, turn left for 10 seconds, repeat for
ever. The feedback path allows the network to drift though this pattern
over and over for every. So no matter what is happening in the
environment, the global feedback allows the system to learn behavior
sequences which are independent of the external environment (like walking).
The normal behavior selection (aka perception, aka attention focus) system,
can then trigger the start of these different cyclic patterns based on the
different environmental conditions, and likewise, abort, or trigger
alternate patterns based on different environmental conditions. Because
it's got inputs from the external environment to use, as well as inputs
from it's internal behaviors, it can learn to focus it's attention on
either depending on what works best. And of course, because it's a huge
parallel network, different things are happening in different parts of the
network at the same time. The right arm can be reacting mostly to signals
flowing in the eyes, while the legs are focused mostly on what the legs are
doing which allows them to produce a cyclic walking motion.
This approach of using another set of inputs to monitor what the systems
outputs are doing is one I strongly suspect is what the brain is doing in
the motor cortex. Though I've never seen a book on the brain explain the
motor cortex that way, I suspect in fact that's exactly how it works and
why it's there. It's not the "output" half of the cortex as much as it is
yet another sensory cortex which is assigned the task of sensing the
outputs signals produced by the entire cortex.
This type of network architecture I believe has all the basic features
needed to produce all human behaviors. That of course is a lot of
speculation, but yet, it seems reasonable to me.
It's a bit odd, but yet simple at the same time, because of the fact that
all human behavior this network can learn has to be trained by a single
reward signal. It has to be structured in a way to make that possible.
It has a lot in common with Brook's subsumption architecture - but the big
difference is that it's a learning network, instead of being hard-coded by
a an intelligent programmer. It has to be structured so that for every
logic function it needs to implement (learn) to create a given intelligent
human behavior, it must have the power to learn that structure through
reinforcement. This basic architecture seems flexible enough to explain how
that can happen.
I've got networks that can learn some very simple tasks using this type of
approach. But what I don't have working correctly is this all important
unsupervised learning function which allows the network to self-organize
into a high quality feature extractor using the temporal clues in the data.
Like I said above, if a rat can't first learn to recognize a flashing
light, it has no chance of learning how to respond to it. And that's where
my networks currently are stuck. They can't learn to recognize a flashing
light! And if they can't learn that correctly, the fact that reinforcement
learning is working to some extent isn't very exciting at all.
This type of approach however is the type of approach which I believe is
needed to make machines act like humans. It will also make them conscious,
but that's a debate for another thread....
It won't make them conscious because it has some "magic feature" needed to
be conscious, it will make them conscious because there is no magic feature
needed to make a machine conscious. Consciousness is a myth - it's an
illusion that humans like to believe in that's nothing like what so many
people think it is. A machine that acts like a human is conscious, and
this type simple learning network, when implemented correctly (big IF
there), I believe will act like a human.
Does that give you a better idea of where I'm coming from and what I
believe in?
Josip Almasi
> cu...@kcwc.com (Curt Welch) wrote:
>> "Isaac" <gro...@sonic.net> wrote:
>>> "Curt Welch" <cu...@kcwc.com> wrote in message
>>> news:20081119002253.348$D...@newsreader.com...
>
>>> I disagree; e.g., how is perceiving the sound of a drum to be a drum a
>>> "problem of behavior selection"?
>> Again, let me explain my approach from the bottom up in a different
>> message and I think you will better understand what I'm getting at there.
>
> Ok, so let me give a bit more of an overview of the type of networks I've
> been playing with and how I think this approach is going to solve AI.
Well thats quite a bit more:)
Just a note regarding drum and behaviour selection.
I don't really remember where I got it from, but here you are: people
respond to music in different fashion. Some people get activity in
motoric, and some in speech centers. So first tend to dance to music,
and second tend to sing/play with music.
So it might be routing issue after all;)
And if you look at it in terms of energy preservation, it's easy to
grasp - senses produce some electricity, which can't just disappear, it
has to go somewhere.
Regards...
Curt Welch
> "Isaac" <gro...@sonic.net> writes:
> >"Neil W Rickert" <ricke...@cs.niu.edu> wrote in message
>
> >> I would suggest that J.J. Gibson's direct perception is more
> >> plausible than Lehar's version of representationalism.
>
> >Could you briefly summarize the Thesis of J.J. Gibson's direct
> >perception?
>
> Gibson's view is that the perceptual system contains transducers
> that are tuned to the invariants of various objects, and that with
> the use of these transducers the objects are directly recognized
> (identified) without and intermediate representation and without
> any computation. Presumably learning involves the construction
> and tuning of the transducers.
This sort of logic fits my view of what is happening in the brain as well.
Though instead of calling them "transducers" I just talk about them as
pattern recognizers, "nodes" in my ANNs, or "neurons".
At the same time, if you arrange these "transducers" into a hierarchical
network, you get both intermediate representations in the hierarchy and
"direct recognition" at the same time.
Maybe something is making Gibson believe the "transducers" shouldn't be
arranged in a hierarchy and as such won't create intermediate
representations but I have no clue why he would want to even suggest that
since we know for a fact that at least parts of the cortex are very much
arranged into a hierarchy (overlapping hierarchies as networks actually) so
to suggest it doesn't work that way strikes me as odd.
The only reason I would see for someone suggesting that is the illusion
that we can't easily sense such a hierarchy at work in our own mind so by
looking into his own mind and trying to guess what he's looking at, he's
making the claim that it there must not be intermediate representations.
> When at the supermarket, there is a scanner that reads the bar codes
> on the items I am buying. The way the bar code scanner works seems
> to better fit Gibson's account than Lehar's representational account.
> There are good pragmatic reasons why the bar code scanner works
> that way, and the same kind of pragmatic considerations are likely
> to have been in play in what made it through biological evolution.
I don't understand what Lehar's representational account is, but all
electronic systems are completely filled with layers and layers of
representations while at the same time, being direct recognition systems.
The difference has nothing to do with the hardware - it's a difference only
in how one chooses to talk about the hardware.
Neil W Rickert
> [asb2]now you have me completely confused. It is impossible to generate an
>error signal (i.e., "lower the tension") without comparing against some
>model of the expected or desired event/result.
I would be interested in how you think auto-focus cameras work.
In particular, what internal model are they keeping and how is the
comparison done to generate the error signal that is used to adjust
the focus.
Isaac
<forbi...@msn.com> wrote in message
news:37f449c3-1ca1-4bc5...@a29g2000pra.googlegroups.com...
>I have recieve the paper and read about 11 pages of it
> and skimmed the rest. It's quite dense. I won't be able
> to get a good feel for several days because I come home
> quite tired.
thanks for your interest. I look forward to your feedback.
>
> Based upon the conculsion I wonder why Stevan
> Harnad's Total Turing Test wasn't mentioned.
Dreyfus is not really saying that AI is impossible, but if it happens that
it will have to be embodied and have a flat, non-representational
architecture. So, it seems like the Turing test and its derivatives is not
something he needs to reference at this point.
Yes, "his way" is what I question in this series of critiques. He does a
lot of hand waving and when he does cite technical details and/or come to
broad conclusions I find that he does without merit.
>Still, digital systems
> in digital worlds are subject to latching where analog systems
> in analog worlds would not. As long as the designed system
> is representational, even if of the system being modelled, it
> is GOFAI all the way down.
>
> I believe your sentence parsing for the sentence starting page 11
> line 33 is incorrect in that Dryfus is attributing a conclusion to
> Chalmers, Clark, and Wheeler rather than making one himself.
> See the next paragraph to see why I believe so.
either way he bases his many conclusions later in the paper on all the
various philosophers he quotes, so to me Dreyfus owns them no matter who is
the source.
Cheers,
Ariel-
jillar...@webtv.net
First, I understand that the "Heideggerian AI failure" means that
"being" in AI is what is failing.
Second, notice how 20th century engineers hide behind the false
philosophical issue of "consciousness" whether they are for it, as Ray
Kurzweil is, or against it, as the pro-illusionists are.
The Heideggarian issue is "being". Start here: Do YOU exist?
How one answers that is something of a "phenomenon" that Heidegger
embodied as a "primordial" experience. Heidegger's paradigm of
"phenomenon" came from Plato, something that Ray Kurzweil has
misinterpretated as "consciousness". Plato never equated a primordial
"form" with "consciousness". A form would be the first and its next-best
"re-publicated" formation. The idea of interpretative circuitry is found
in Plato's allegory of the cave. "Phenomenon" has three meanings: 1]
original, 2] copy, 3] scientific object. How would a copy respond as a
scientific object that it exists as something original?
As for critique #1: you have not answered Heidegger. You have switch the
topic to "consciousness" and the very silly notion [in
applied-scientific terms] of the "unconscious". I'm not intending to be
contentious. Just keeping the track on Heidegger and why MIT attendants
and other engineers would consider an Heideggerian AI model?
regards,
Vjillar
forbi...@msn.com
> <forbisga...@msn.com> wrote in message
> > I believe your sentence parsing for the sentence starting page 11
> > line 33 is incorrect in that Dryfus is attributing a conclusion to
> > Chalmers, Clark, and Wheeler rather than making one himself.
> > See the next paragraph to see why I believe so.
>
> either way he bases his many conclusions later in the paper on all the
> various philosophers he quotes, so to me Dreyfus owns them no matter who is
> the source.
Why would he own something he is critiquing?
The passage I was referencing is:
Heidegger’s important insight is not that, when we solve
problems, we sometimes make use of representational
equipment outside our bodies, but that being-in-the-world
is more basic than thinking and solving problems;that it is
not representational at all. That is, when we are coping at
our best, we are drawn in by solicitations and respond
directly to them, so that the distinction between us and
our equipment--between inner and outer—vanishes.
As Heidegger sums it up:
I live in the understanding of writing, illuminating, going-in-
and-out, and the like. More precisely: as Dasein I am --
in speaking, going, and understanding -- an act of
understanding dealing-with. My being in the world is nothing
other than this already-operating-with-understanding in this
mode of being.
So in the paragraph to which you responded:
My comment:
"Assuming that by "thinking" you mean conscious thought,
I cannot see how thinking is a bridge that necessarily follows
from memories/beliefs not being solely inner entities. It
seems to me that inner and outer representations can be
bridged without thought.
You are responding as if Dryfus is supporting a position he's
just reporting. Even in the paragraph following he isn't taking
a position but clarifing Heidegger's position and describing how
Wheeler either doesn't understand it or is misrepresenting it.
In another thread I allude to the symbolic/subsymbolic nature
of things. I'll go into a bit more detail. Suppose I have an
electronic device we call a nand gate. Two leads are identified
as inputs A and B and one lead is identified as output C.
C will have about 5 volts on it as long as A and B do not
have about 5 volts on both of them. Here's the table without
going into a bunch of detail.
A B C
~0v ~0v ~5v
~5v ~0v ~5v
~0v ~5v ~5v
~5v ~5v ~0v
Normally 0v is represented as 0 and power voltage as
1. This works for positive and negative power so ~5v
could as easily have been ~-5v. Heck, just saying 5v
shows my age.
Here's the thing...
Not only is the assignation of false to 0 and true to 1
not part of physical reality but neither is consistent
assignation of false and true to the same voltage at
all leads. Here's another table where F represents
false and T represents true:
A B C
F T T
T T T
F F T
T F F
That's the same table as before but with different
logical assignments. When we use such a device
it's purpose isn't set in stone and we are not relying
upon its representation of anything. There's nothing
representational in the world to bridge.
Again:
Heidegger’s important insight is not that, when we solve
problems, we sometimes make use of representational
equipment outside our bodies, but that being-in-the-world
is more basic than thinking and solving problems;that it is
not representational at all. That is, when we are coping at
our best, we are drawn in by solicitations and respond
directly to them, so that the distinction between us and
our equipment--between inner and outer—vanishes.
Contrary to what I said before Dryfus is in fact telling us
his position on the matter when he uses the words "...
important insight..". Dryfus is saying that Wheeler was
wrong and his interpretation of Heidegger was wrong.
Some have said that the enemy of my enemy is my friend.
I was about to be drawn into this position by saying that
when you respond to Weeler's position you support Dryfus
but that isn't so. Dryfus follows the Heiddeger line where
you respond: "It seems to me that inner and outer
representations can be bridged without thought." and this
falls well off the mark concerning their position on the
matter.
Dryfus is talking about being in the world and coping with
it intelligently.
Page 32 line 22:
It would be satisfying if we could now conclude that, with
the help of Merleau-Ponty and Walter Freeman, we can
fix what is wrong with current allegedly Heideggerian AI
by making it more Heideggerian.
Note the word "allegedly". I don't know how you can correctly
assert, "either way he bases his many conclusions later in
the paper on all the various philosophers he quotes, so to me
Dreyfus owns them no matter who is the source." when he is
basing his conclusion on a denail of their positions being
Heideggerian.
Curt Welch
> > I'm an engineer, not a philosopher.
>
> >As such, nearly everything you write
> > strikes me as silly and odd and misguided.
>
> I am an engineer, scientist, philosopher, and roboticist. Of course, the
> problem does not reside strictly in any one dicipline or skill set, so I
> am not surprised that an implementation oriented thinker will find the
> abstractions too obtuse for utility.
Then you are one of the lucky rare ones! :)
Yes, anyone who can understand and combine all the disciplines together
will have a clear advantage.
> I disagree. Reverse engineering will not solve the problem and may
> actually lead to many dead ends. It will take a new theory and
> philosophy to do it. Think of it like trying to emperically come up with
> QED or Relativity w/o any new theory or philosophy of physics.
Yes, you must have the theory to guide you. That's the hypothesis in
science. But at the same time, the theory must be distilled from empirical
evidence or else it's not very likely to do much good. Many of the
philosophers who don't do a good job of crossing over between fields fails
to do that in my view.
> > It's not a problem which can be solved by pure
> > philosophy.
> >
> True, but you can't just do it bottom up either. You can miss the big
> picture, which philosophy can shed light on.
Right. I spend endless hours philosophizing about the big picture so I
know where to head. But you have to build from the bottom up - aka do
empirical experiments to verify or reject your philosophizing.
> > For example, you speak of this "unity between the mind and the world".
> > What exactly is the "mind" and the "world"?
>
> I did not say this. If you read my intro, I was quoting from Dryfus'
> paper.
Sorry, your quoting style confused me I guess. People sometimes quote one
paragraph and then write their response starting with the second paragraph.
I don't think I was sure what was the quote and what might have been you
talking.
I like to indent any material I quote to make it obvious.
> >You can't resolve this sort of
> > question just by talking about such things. Words are defined by their
> > connection to empirical evidence and without empirical evidence, the
> > words are basically meaningless - or at minimal, available for use in
> > endless pointless debates and redefinition based on usage alone.
>
> for sure symantics can lead to circular definitions, but tossing out
> anything not empiracle is "throughing the baby out with the bath water";
> that is, you toss out powerful abstractions that bridge large gaps
> empirical evidence.
Yeah, you can start with any idea no matter where it came from, but if you
can't verify it using empirical data, then the odds of being useful become
highly reduced.
> > The problem we run into here is that without a concrete definition of
> > how the brain works and what the mind is, we can't make any real
> > progress on the types of issues you are touching on here. How can we
> > make any progress
> > debating the nature of the connection between the "mind" and the
> > "world" when we can't agree what the mind is? And if we can't agree
> > what the mind is, we can't really agree on anything it creates - like
> > it's view of the world - which is the foundation of what the word
> > "world" is referring to.
> >
> > You can't resolve any of these questions until you can first resolve
> > fundamental questions such as the mind body problem and consciousness
> > in general.
> >
>
> Well, we have to talk about the trinity or we'd get no where, but I agree
> that any usage of those words must be very tentative and cannot lead to
> sweeping conclusions w/o a scientific definition of each, which I say
> would require a theory of mind (not connecting a million data points).
Well, I think mind and consciousness in fact only slow people down and that
it would be solved just fine if the ideas were never mentioned. The entire
concept grows from an illusion which makes people think there is something
happening in them which simply isn't there at all.
The only point to studying mind and consciousness is to understand the
illusion. However, almost no one seems to understands it's an illusion and
as such, waste endless hours trying to understand something which is a
ghost instead of ignoring it like they should be doing. It's just this
illusion which gets the philosophers into such much trouble when try to
figure out the mysteries of the brain without using empirical data.
> > I have my answers to these questions, but my answers are not shared or
> > agreed on by any sort of majority of society so my foundational beliefs
> > can't be used as any sort of proof of what is right. It call comes
> > back to
> > the requirement that we produce empirical data to back up our beliefs.
> > And
> > for this subject, that means we have to solve AI, and solve all the
> > mysteries of the human brain. Once we have that hard empirical science
> > work finished, then we will have the knowledge needed, to resolve the
> > sort of philosophical debates you bring up here. Until then, endless
> > debate about what "merging mind and world" might mean, is nearly
> > pointless in my view.
> >
> > Having said all that, I'll give you my view of all this, and the
> > answers to
> > your questions as best as I can figure out.
> >
> > I'm a strict materialist or physicalist. I believe the brain is doing
> > nothing more than performing a fairly straight forward signal
> > processing function which is mapping sensory input data flows into
> > effector output data flows. There is nothing else there happening, and
> > nothing else that needs to be explained in terms of what the "mind" is
> > or what "consciousness" is.
>
> I don't think you can call anything as chaotic as the brain doing
> anything "straight forward". The Earth's weather is infinitely more
> straitforward than the humand mind/brain and we cannot model it worth a
> damn even with all the most powerful computers in the world.
Only the behavior is chaotic and too confusing to understand. The
underlying principles which create the chaos is what I believe to be very
straight forward. I fyou can understand the straight forward underlying
principles, then you can understand the chaos.
Random number generators are also chaotic and too confusing to understand.
But yet, no one thinks twice about them as being some big mystery in the
universe because the underlying principle that causes the chaos is simple
and straight forward. I think the brain is exactly the same sort of thing.
> >The mind and consciousness is not something separate
> > from the brain, it simply is the brain and what the brain is doing.
> >
> > It's often suggested that humans have a property of "consciousness"
> > which doesn't exist in computers or maybe insects (based on the use of
> > "insect" above). I see that idea as totally unsupported by the facts.
> > It's nothing more than a popular myth - and a perfect example of the
> > nonsense that is constantly batted around in these mostly pointless
> > physiological debates.
> >
> > There is no major function which exists in the human brain which
> > doesn't already exists in our computers and our robots which are
> > already acting as autonomous agents interacting with their environment.
> > The only difference between humans and robots, is that humans currently
> > have a more advanced signal processing system - not one which is
> > substantially different in any significant way - just one which is
> > better mostly by measures of degree, and not measures of kind.
>
> I don't think you could be farther away from the truth. The brain
> computes in ways that is so different (an often oposite) of how our
> signal processing works that it is in another universe by comparison.
> For example, the core of the brain's sensory processing seems to be a
> kind of synethstesia based system, which is exactly what all engineers
> would avoid like the plague. I could go on and on with counter examples.
Sure. But not all of us avoid that type of design like the plague. All
reinforcement learning programs and neural networks work that way. It's
the general field of computer learning you are describing.
Though as a programmer, you have no hope of being able to program a neural
network to recognize faces by hand adjusting the weights. That's why we
don't program that way. We develop learning algorithms to do the
programing for us.
Learning to program learning algorithms is tricky business and it's exactly
why solving AI is taking so long. You don't get to just program in the
behavior you see.
If I gave you a long sequence of random numbers (a whole book full) how
long do you think it would take you to come up with the code that produced
it? It's a reverse engineering task that is as hard as breaking a cipher.
Solving AI is the same class of problem.
I have no doubt that the brain is using simple and straight forward
learning techniques to shape the chaotic behavior of a network of
communicating neurons. TO solve AI, we have to figure out what those
simple and straight forward learning techniques are.
> hebbian learning was known since the '50's but that has not lead to
> anything practical because it may necessary but not sufficient.
Hebbian learning is hardly a complex enough concept to even be called a
learning algorithms. It's high level hand waving (which as John will tell
you I do plenty of on my own). It's like the first step down a long path
of developing learning algorithms. It's the idea "maybe there are simple
rules to explain how neurons learn, and if we we use these rules in a large
network - the network will produce intelligence".
Yes, I believe in that idea, but it's not the answer, it's little more than
first step at trying to find the answer.
> For
> example, hebbian learning does not even begin to solve the frame problem.
> Since this is so strait forward, how do you propose reinforcement
> training (i.e., Pavlov's dog) can be used to robustly deal with the frame
> problem?
The little I've seen of the frame problem, the question in general it seems
silly and absurd - it seems to be a problem created by asking the wrong
question and making the wrong assumptions.
However, this might be a good one to debate with you to see where it leads.
I've been slow in reading and responding to these questions, and a post or
two I wrote last week about my views touched on the Frame problem. I don't
know if you wrote the above before I after I posted that. Maybe I should
start another thread so we can chew on the Frame problem a bit....
> > From this perceptive, let me jump in and debate the words you had issue
> > with:
> >
> > we are drawn in by solicitations and respond
> > directly to them, so that the distinction between us and our
> > equipment--between inner and outer-vanishes
> >
> > I think at the lowest levels of what is happening in the brain, it is
> > obvious that the brain is simply reacting to what is happening in the
> > environment - that is what the brain is doing by definition in my view.
> > We
> > simply "respond directly to them". That is all we every do.
>
> really? So, being an engineer you will know that "reactions" to input is
> just another way of saying that you have a control system.
Sure.
> However, any
> control system needs a model to determine the proper control surface for
> the input landscape; that is, model building.
No it doesn't.
> Thus, the brian is about
> building useful models of the environment via sensory synergy. In this
> way, I completely disagree with your assertion that "brain is simply
> reacting to what is happening in the environment "
We build models to _explain_ the way the brain reacts. In fact, the brain
doesn't have to build models, it only has to react. Or, from another
direction, the way it reacts, is what you are calling "the model".
Look at Brooks' subsumption systems. Where are the models in those
designs? This was exactly the point he was making in that approach.
Complex goal driven behaviors that seem to be using a "model" can be
implemented as priority driven reaction systems.
> Well, Dreyfus disagrees with you on this point. He says there is no
> representation of a dog in the brain. How do you argue against that?
I think representation is in the eye of the beholder. :) That is, when we
look at hardware, we can describe it as having a representation or we can
choose to believe it doesn't.
Any causality chain that exists in the universe is an example of
representation and you can be dead sure the brain is full of causality
chains.
If a red ball hits a blue ball and makes the blue ball change direction,
the new direction of the blue ball is a representation of the old direction
of the red ball. All causality chains create representations. Whether you
choose to call it that when you are talking about the physics of what is
happening is totally a choice of your frame of reference.
I've not read Drefus so I can't comment directly on what I think his view
might be.
As I've explained in other messages I've posted now, I think the way the
brain works, and the way we need to build hardware to duplicate human
behavior, is to create networks which translate sensory data into a large
set of signals which represent micro-features of the current environment.
All those signals collectively represent the systems view of the current
state of the environment.
I think the brain very much has a representation of dog in the brain. It's
almost absurd to say it doesn't because if it didn't, we couldn't think
about a dog (assuming you are a materialist like I am believes that thought
is just brain behavior).
Now, if you want to argue that there is no grandmother cell, that's a very
different argument. That's not an argument against representation, it's an
argument about the form of the representation. I would agree there
probably is no grandmother cell. I believe there are lots of micro-feature
cells and that some general cluster or cloud of micro features is our
brain's representation of a high level concept like "grandmother". If
that's the argument Dreyfus was making about the dog then I would agree
with him.
> >We can look right at it, and have no clue what it is
> > we are looking at.
>
> All the evidence I am aware of re the brain is that such concepts are not
> located in any one place which you can damage to lose only the
> recognition of a dog. BTW, this is another example of how the brain is
> radically different than our computing systems.
Right, simply because a high level generic concept like dog represented by
a single English word does not exist in our brain as single micro-feature.
It exists as a large and loosely defined cloud of micro features - a
collection that doesn't have any absolutes but rather just a large set of
probability. If you damage part of that cloud you make it harder for the
brain to use that concept, but the concept of dog doesn't die completely
unless you can damage most the micro-features that define it - and that
might damage such a large part of the brain so as to make the wipe out a
lot more than just the "dog" concept - you might kill the brain's ability
to use language in general.
> > At the same time, if you stimulate the correct parts of
> > a brain, it's mostly likely that the person would report they were
> > "seeing a dog" when there was no dog there.
>
> There is no research ever showing that this is possible. Please cite the
> research that supports your belief. I only know of music being able to
> be stimulated to be heard in the brain.
I have no such examples. What me, use empirical evidence! :)
What I suspect is that you would have to stimulate enough of the correct
micro-features (neurons) that represented the idea of "dog" in the brain
before the person would be stimulated into thinking the dog concept.
Without better brain scanning and stimulating tools we couldn't locate that
potentially well scattered set of neurons nor activate them.
However, what we can do, is stimulate sections of the brain and ask the
patient if they sense _something_. The answer, as far as I know, is that
the patient does sense things when they are stimulated in different parts
of the neo cortex. What we can't do, is accurately know what part to sense
to to create what type of sensation or thought.
BTW, the empirical evidence I use is not brain research, it's computer
science research into the behavior of algorithms. All my talk about brains
is mostly speculation based on how I think it must be working based on what
I know about making computers work. Keep that in mind when I say things
like "the brain does....".
> > So what is it we are actually
> > "seeing". Is it the dog we are responding to when we say we see a dog,
> > or is the neural activity in one part of the brain which other part of
> > the brain is responding to by producing the words "I see a dog"?
> >
>
> I think we should stay away from consciousness in this discussion or else
> we will get no where by forking out to too many infinities.
:)
> > It can be argued that what we actually respond to is not the physical
> > dog out in the word, but that we are responding to the brain activity.
>
> Of course, the model of the dog.
Yes, I can call this network of micro features a model, or I can call it
"the way the brain reacts to the environment". It's the same thing either
way to me. I think it's important to understand it from both views.
> I believe he means our brain circuits engage phenominon by melding with
> it and becoming a mirror image such that the two are not seprable, thus
> no representations of the object in the brain just a bunch of organically
> melded dominoes that hit one to another like "reality" would.
I really can't grasp what you are suggesting there.
If the brain is "dominoes that hit one to another like reality would" then
that to me is a dead on definition of what a model is. It's a
representation of the object using dominoes.
Also, as I said above, I think all causality chains that exist in the
universe are representation systems. Representation is just a way of
describing causality. For example, the height of the mercury in the
thermometer is a representation of the average kinetic energy of the air.
The location of the dial on the pressure gauge is a representation of the
pressure in the pipe. The angle of the wind sock is a representation of
the air speed. The foot print in the mud is a presentation of the deer
that walked past an hour ago. The mud puddle is a representation of the
rain that happened an hour ago. The electron charges in the flash memory
card of the camera is a representation of the person that was in front of
the camera the last time the exposure button was pressed.
> I believe he is saying that phenomenon is internalized w/o distinctions;
> i.e., you become the phenomenon. As opossed to you making a model of the
> object as a seeprate token to use in your brain system to plan your
> actions.
Well that wording has a little bit I can relate to. That is, the way our
brain reacts as a whole is the representation instead of something in the
brain creating a representation and them manipulating it. That is, how we
react to it is the representation instead of the representation being
something we manipulate (as we would manipulate a model with our hands).
That is, there is not one part of the brain which is the "us" and another
part of the brain which is "the model". The brain is us and how we react
to the world is the model. We are the model.
In this sense, I think I agree with what he was saying - if that is in fact
what he was thinking.
> In his paper, he considers chaotic neural networks as being more at "one
> with the world" than classic AI's more (fuzzy) rule based systems.
Well, with my view of the network being a set of micro features that
represent the current state of the environment, I think we are in sync with
the idea of being "one with the world".
casey
> The only point to studying mind and consciousness is
> to understand the illusion.
Driving along the road you may have the illusion of
shimmering water in the distance. That it is water is
an illusion, that you experience it is not an illusion.
You do not explain the illusion, you deny it, claiming
it is a conditioned belief we have been trained to see
the water but it doesn't exist. You claim there is
nothing to explain. But illusions can be explained
Curt. An illusion is not a nothing. And it is not
simply a conditioned belief. Or are you going to
say a mirage doesn't exist we only think we are
seeing a mirage because we were trained to think
we were seeing a mirage?
To me that is silly beyond belief. Yes the illusion
is that it is water. Yes the illusion was that the
experiences we call mental events was a soul. But a
false explanation of the illusion being experienced
doesn't make the illusion a result of conditioning
or training. The belief in the soul was an explanation
of something not a belief in something that didn't
exist. We do have experiences.
I explain it as two views of the same thing.
It is an illusion a machine will have without the
need for any training or conditioning.
JC
Curt Welch
> On Nov 26, 10:59=A0am, c...@kcwc.com (Curt Welch) wrote:
>
> > The only point to studying mind and consciousness is
> > to understand the illusion.
>
> Driving along the road you may have the illusion of
> shimmering water in the distance. That it is water is
> an illusion, that you experience it is not an illusion.
> You do not explain the illusion, you deny it, claiming
> it is a conditioned belief we have been trained to see
> the water but it doesn't exist. You claim there is
> nothing to explain. But illusions can be explained
> Curt. An illusion is not a nothing. And it is not
> simply a conditioned belief. Or are you going to
> say a mirage doesn't exist we only think we are
> seeing a mirage because we were trained to think
> we were seeing a mirage?
>
> To me that is silly beyond belief.
Give it up John.
I've explained this to you at LEAST 20 times and you still don't understand
any of it. You probably never will.
But, just one more time...
The ILLUSION OF CONSCIOUSNESS (that I talk about all the time), is that
private mental events (our thoughts, aka subjective experence) don't seem
to be neurons firing, even though they are.
I'm am DENYING NOTHING!!!! I know we have private mental events, and that
these form our subjective experience of the world we live in.
I also have never, ever, even once, said there is nothing to explain.
There is something here to explain. What we have to explain is why our
subjective experience doesn't seem to be neurons firing when they are.
But, I've explain exactly why this happens many times to you as well.
But yet, you don't understand any of it. You don't understand the
explanation. You don't even understand that it's been explained to you.
And not only do you fail to understand, you accuse _ME_ of being suffering
an illusion when I tell you that I'm able to sense my neurons fire. Again,
right over your head without any understanding of what I've explained to
you 20 times.
We all experience this same illusion for a very simple reason. We don't
see our private subjective experience as being the neuron's firing because
we have never been exposed to data which allows us to correlate neuron's
firing with subjective experience. The brain builds links between
different stimulus signals based on correlations and because there haven't
been any for most of us because we have never been able to see our neuron's
fire while experiencing the subjective experience at the same time.
Let's call that 21 times I've explained it to you now. But, like all times
in the past, I don't expect you to get anything out of this time either.
The problem here is that you think the illusion is real. You think that
not only do we have neurons firing, but that they somehow _create_ this
other thing, called subjective experience. And as such, you think we need
to explain how the "other thing" gets created. And your current favorite
answer to say it's produced by some brain process we don't understand. But
there is no such thing as "subjective experience" which is _separate_from_
brain activity, so in that sense, there is nothing to explain. All there
is to explain, is the like brain activity.
The basic nature of the process which creates the illusion I do understand,
and if you want to understand, all you have to do is keep reading what I
wrote above, until you understand what those words mean. All the mysteries
of consciousness are explained in the words above if you care to read them
and understand what they imply. There are a TON of people in this world
that are simply unable to understand those words - so you are not in any
sense alone.
casey
> Give it up John.
I have given it up Curt. I will let you continue
with your crusade without comment.
Curt Welch
> "Curt Welch" <cu...@kcwc.com> wrote in message
> news:20081116234549.142$i...@newsreader.com...
> hebbian learning was known since the '50's but that has not lead to
> anything practical because it may necessary but not sufficient. For
> example, hebbian learning does not even begin to solve the frame problem.
> Since this is so strait forward, how do you propose reinforcement
> training (i.e., Pavlov's dog) can be used to robustly deal with the frame
> problem?
I wanted to just followup and see if we could get a discussion of the Frame
problem going since it seems to be a recurring interest to you.
BTW, Pavlov's dog experiment is an example of classical condition, not
operant conditioning. But that's not important.
I've never understand why people think the Frame problem is even a real
problem, so you are going to have to give me some examples of real world
things that humans do that are examples that demonstrate our ability to
solve the Frame problem but which you don't understand how AI hardware will
solve it, and I'll be happy to explain how I think AI hardware will solve
it.
Using this is a starting point:
http://en.wikipedia.org/wiki/Frame_problem
It talks about the frame problem in AI starting as a logic problem. That
form of the problem was created simply by trying to use logic as system of
representing information about the environment. It's a problem crated by
picking the completely wrong implementation paradigm for trying to solve AI
and doesn't apply in that form to any other issue in AI. As such, it's
just not important.
As I've explained in other messages, I believe the foundation of
intelligent behavior is a reinforcement trained connectionist network and
since it's not using that type of logic to try and represent the state of
the environment, the frame problem just doesn't apply. If you think it
does, please explain how it does.
But, the wikipedia article also makes reference to the broader form of the
idea that shows up in philosophy which has something to do with "updating
beliefs" as to how the environment changes in response to an action.
I guess you issues are more along this general question somehow???
I believe AI will be created by a machine that learns by experience. It
learns what reactions to a given state of the environment works, and what
doesn't work.
Such as system simply needs to learn how the environment changes by
watching it change. It's only "beliefs" are based on the assumption that
if the environment changed the same way the past 10 times in response to an
action, it can be expected to change the same way with a high probability
this time.
So where's the Frame problem in this? A system which has the power to
"remember" how the environment has responded in the past and works on the
basis that whatever probabilities guided the change in the past are likely
to guide change in the future doesn't seem to have a frame problem to me.
So what is the Frame problem to you and why do you see it as something
which is so hard to solve?