From: Franklin Wayne Poley (culturex@vcn.bc.ca)
Date: Wed Sep 06 2000 - 17:13:43 MDT
What I am trying to determine in this dialogue with Professor Minsky is
the extent to which a profile of real, human intelligence can be matched
by artificial, humanoid intelligence. Do we have the knowledge of how to
do so now? Is it 'just' a matter of applying more labor and money to the
knowledge we have or do we need to invent some new technologies? If we
need to invent new technologies, what are they? Can anyone answer these
questions?
FWP
---------- Forwarded message ----------
Date: Tue, 5 Sep 2000 17:01:48 -0700 (PDT)
From: Franklin Wayne Poley <fwpoley@vcn.bc.ca>
Reply-To: EDTV-Robotics-State-Of-The-Art@egroups.com
To: EDTV-Robotics-State-Of-The-Art@egroups.com
Cc: minsky@media.mit.edu, brooks@ai.mit.edu, kw@cyber.reading.ac.uk,
hans.moravec@rover.ri.cmu.edu, raymond@kurzweiltech.com,
conventions@mures.com.au, jfe@helpmate.com, don@jnd.org,
elubofsk@cognex.com, humanoids@usc.edu
Subject: [EDTV-Robotics-State-Of-The-Art] Real Human Mental Ability vs.
Artificial Humanoid Mental Ability: Profiles and the "Human Equivalency"
Question
On Sun, 3 Sep 2000, Franklin Wayne Poley wrote:
> > Date: Tue, 29 Aug 2000 21:46:22 -0400
> > From: Marvin Minsky <minsky@media.mit.edu>
> > To: Franklin Wayne Poley <fwp@vcn.bc.ca>,
> > EDTV-Robotics-State-Of-The-Art@egroups.com, brooks@ai.mit.edu,
> > kw@cyber.reading.ac.uk, hans.moravec@rover.ri.cmu.edu,
> > raymond@kurzweiltech.com, conventions@mures.com.au, jfe@helpmate.com,
> > don@jnd.org
> > Subject: Re: Can HomeMate Surpass Human Equivalency in Measured
> > Intelligence?
> >
> > Franklin,
> >
> > >"The only barrier today to a useful humanoid robot is money. We have
> > the
> >
> > >sensory perception, we have voice recognition and voice synthesis; we
> > have
> >
> > >sufficient computer power. We have all the tools." (p. 88).
> >
> >
> > Engelberger is flimflamming here. We have only some limited,
> > superficial, input-output systems, but vitually none of the reasoning
> > tools that are typical of a five-year-old.
Real human intelligence is simulated by artificial humanoid intelligence.
Those 19 factors of primary mental ability referred to below give us a
pretty good estimate of 'g', general intelligence, along with some useful
details. If anyone wants to study a practical application of this they can
get a test kit for Cattell's CAB (Comprehensive Abilities Battery) which I
have used in clinical practice. The factors are a composite of abilities
so my "common sense" taxonomy of Visualization, Reasoning, Memory and
Verbal is largely to facilitate inter-disciplinary communication.
Perhaps AI cannot simulate the reasoning tools of a five-year-old
because we are looking for a simulation of reasoning which is too
"esoteric". In other words we are bringing "mentalism" back into
psychology which is what Watson's 1913 Psych Review paper tried to
eliminate. I think Watson's position is valid for practical purposes like
studying real human intelligence and artificial humanoid intelligence.
Four of the 19 factors are readily categorized as Reasoning
factors: Numerical Facility (N); Deduction (D); Induction (I); General
Reasoning (R). My question (and it is a question) has to do with whether
all of the field of arithmetic/logic/mathematics can be programmed
(assuming enough worker hours for such a mega-project and a sufficiently
powerful computer). Can anyone select any exercise in
arithmetic/logic/mathematics, put it on the internet and say, "Here is an
exercise which cannot be programmed?"
(a) yes
(b) no
(c) is for esoterics. Here one can say, "Yes but....". The but part is for
"but this is just rote; it is not real reasoning". No it's not real human
reasoning. It is artificial humanoid reasoning. Machines don't do their
reasoning operations the same way as humans. Why should they? Moreover, is
real human reasoning all that superior? I use very rote/mechanical methods
when I am solving Factor N tasks. I do the same if I am solving a Factor I
problem, eg "What comes next in the series: 1 3 5 7 9 _ ?" It becomes a
moot point as to whether I am REALLY thinking with my "real human
intelligence" but the machine is not with its "artificial humanoid
intelligence".
So it comes back to the basic question as to whether anyone can give a
specific example in arithmetic/logic/mathematics in which the artificial
humanoid intelligence cannot simulate real human intelligence sufficiently
well (like an arithmetic program) to meet "human equivalency" criteria.
Some additional comments on the 6 "Visualization" factors below.
> Professor Minsky:
> Mr. Engelberger's position is that robotics is not the
> same field as AI and that he is able to say authoritatively as an expert
> (indeed one of the 'founding fathers', with Devol, of modern
> robotics) that robots can now do more than one would expect. However, he
> has cautioned me in another email which I will get to later, re
> overstating "human equivalency". Those words of caution are being heeded
> and I apologize if I gave the impression that he had put human equivalency
> forward as his position. "Human equivalency" is an apt phrase coined by
> Professor Moravec and I am using it for the Robotics-State-Of-The-Art
> program as an interesting, publicly-appealing and unifying thread to tie
> the whole program together (exectedly 1-2 hours). I am not pushing it as a
> matter of what the state-of-the-art is, unless the facts (as revealed by
> experts in interviews with Mr. Ingram) reveal otherwise. Thus it is the
> FACTS which I must ascertain in order to prepare a script which lives up
> to the name for this program. That is the purpose of this EDTV- list.
> I thank you for the examples you have given below. Those are exactly
> the kinds of issues which the program must address, one by one. Let me
> preface my comments on them by saying the obvious which is that indeed AI
> is not robotics and robotics is not nanotechnology and none of the above
> are alife. So there is a great broadening and deepening of the field.
> By analogy I wonder how long it was before William James, the 'founding
> father' of American psychology was unable to claim expertise on all
> branches of scientific psychology? If he were living today, he couldn't
> even come close.
> It is quite an honour for me to be able to correspond with both the
> founding father of robotics and the founding father of AI (yourself and
> John McCarthy) at the same time. I hope we can come to an agreement with
> the Ingram Studio so that both of you will be interviewees. I'd like to
> focus here on human equivalency of the RESULTING PRODUCTS OF AI only. What
> are the FACTS? To explain "resulting products" let me refer to pages 41-43
> of my "Individual Differences" text (Gardner Press, 1976). I list 19
> primary mental abilities from factor analytic studies going back to
> Thurstone. I give brief descriptions of the factors and sample test items.
> Cattell and colleagues figure most prominently in carrying forward this
> work on intelligence testing and multi-factor studies from Spearman to
> Thurstone to Cattell to the next generation like Professor Hakstian at
> UBC and now his students are the current generation. So here we have five
> generations of people who have dedicated their life-time careers to real,
> human intelligence. I say that because by this time it is as unlikely that
> a significant new factor or even test item will be discovered as that a
> new element will be discovered in chemistry.
> Thus those 19 factors cover 90% + of what you would want to deal
> with in measuring G, general intelligence. Now here is the important part
> for Robotics-State-Of-The-Art.
>
> DO THE RESULTS OR PRODUCTS OF THESE TESTS MEET HUMAN EQUIVALENCY CRITERIA
> WHEN SIMULATED BY AI ?
>
> The best example to give first is Factor N (Number Facility) on page 42.
> This is arithmetic ability. Can't calculators meet the criterion of human
> equivalency in arithmetic ability? They can certainly surpass my
> arithmetic ability.
>
> Moreover, it looks to me like the other 18 factors can be simulated as
> well, when it comes to arriving at a result or product as we did with
> Factor N. Does that not mean that AI has achieved human equivalency when
> it is defined this way?
>
> There still is no computer
> > vision system that you could put into an unprepared room and have it
> > recognize books, tables, chails, lamps, cups, and things like that.
> > 'perception' is not enough to be useful.
>
> I have cc'd the Cognex Company which identifies itself on its web site as
> the world leader in machine vision. "Visualization" is an important
> aspect of the 19 factors (along with Verbalization, Memory and Reasoning
> as 'common sense' dimensions). E.R. Davies writes in his 1997 Machine
> Vision text "...there is an important guiding principal: if the eye can do
> it, so can the machine...." (p.14). In what SPECIFIC ways can machine
> vision meet the human equivalency criterion? The East London Burrough of
> Newham uses surveillance cameras to automatically identify known criminals
> on the street. See <http://www.faceit.com> and Raymond Kurzweil writes in
> his 1999 book, that face recognition technology is "...evidently reliable
> enough that the banks are willing to have users walk away with real
> cash." (p. 77). If www.faceit software can identify a face in a crowd can
> it not be modified to identify a book on a table? There may be some reason
> it cannot do so and that is why I ask. If we go to
> <http://www.braintech.com> we find that the "object recognition system" as
> Braintech calls it is sufficiently well developed to handle the "binning
> and sorting process with multiple objects". Can it not pick a cup out of a
> bin of mixed objects? So the broader question of perception/visualization
> makes me wonder what specific examples of tasks are beyond present machine
> capabilities. Perhaps the examples above (book, cup etc.) entail some
> unsolvable features but I don't know that for a fact yet and I am hoping
> Mr. Lubofsky at Cognex can inform us of exactly where the
> machine/software capabilities fall short of human equivalency.
> Quite a lot of machine vision software is now coming onto the market.
> The Davies quote above sums up the overall question and it is another way
> of saying "human equivalency". But it is a question, not a firm
> declaration. The Omron Company at <http://www.qub.ac.uk/ivs> lists its
> F400 system which has "hue, saturation and intensity-based color
> detection." I have to ask how well this system compares to human detection
> of hue, saturation and intensity? Moreover, if it falls short, what could
> be done with more money for a better system? The F400 system has
> "position compensation (which) detects objects regardless of
> orientation". How well can it detect those objects in a room above?
> In testing human mental abilities, a number of visualization tasks
> entail visualizing objects from different perspectives. Like other mental
> abilities one can readily see "real world applications" though mental
> measurements people are quick to point out that the initial idea of
> finding particular profiles of primary mental abilities which would match
> to specific vocations was long ago shown to have little validity. However,
> I used to work with surveyors and make maps in a mine so I am quite aware
> of the parallel between test items here and real world problems. If a
> miner for example could look at a map from one perspective, then quickly
> and accurately visualize the shafts and stopes and drifts (excavated
> regions) from another perspective, we would recognize that as a sign of
> intelligence. I have to expect that a computer program could do this much
> better than a human. If there is any doubt we could ask Dr. Baiden from
> Inco with whom we corresponded earlier. I think he will be an interviewee
> (subject to final approval by Inco).
> And what about depth perception? That policeman's radar tells me that a
> machine has better depth perception than a human (as well as speed
> estimation if one wants to call that a Visualization ability). What else
> can we say about the comparisons here? In what ways can machines measure
> depth better than humans and in what ways are they inferior? Again, is
> there any shortcoming of machine depth perception which more money cannot
> remedy?
> Machine Visualization is obviously better than human when it comes to
> detecting light at very low intensities and very high
> intensities. Machines can detect outside the human spectrum (eg infrared
> and ultraviolet).
> Thus we can carefully put together a profile of Human vs. Humanoid with
> respect to the primary factors of mental ability related to Visualization.
> If we can work this out then I would like to go on to the other problem
> situations presented in your posting, Professor Minsky. And we will look
> at factors in the areas of Reasoning, Verbalization and
> Memory. Immediately we know that machines surpass human equivalency in
> a number of ways. So we need to also know: (1) For what reasons, exactly,
> do they fall short at present? (2) Do we have the know-how, subject to
> large infusions of money, to bring them up to human equivalency?
>
> Thank you again for your help on this matter.
>
> Sincerely-FWP
>
> http://users.uniserve.com/~culturex/Machine-Psychology.htm
Visualization factors among the 19 primary mental abilities are: P,
Perceptual Speed; Vi, Spatial Visualization; S, Spatial Relations, SO,
Spatial Orientation; Cs, Speed of Closure; Cf, Flexibility of Closure.
If we can meet human equivalency for these 6 factors plus the 4 Reasoning
factors we have 10 of the 19 taken care of. The mental abilities profile
of human vs. humanoid is getting interesting.
A key to this is having the best interviewee possible on OBJECT
RECOGNITION SOFTWARE.
Is it now possible to write software so that any object can be recognized
with a cluttered background?
Can the cup and book above be recognized regardless of how cluttered the
field is around them?
Next question: Is there any way in which a machine cannot detect the
distance of an object as well as a human or better? (artificial humanoid
depth perception vs. real human depth perception).
Next question: Can machines detect as many pixels/unit of area as a
human? (another measure of human equivalency).
FWP
To unsubscribe from this group, send an email to:
EDTV-Robotics-State-Of-The-Art-unsubscribe@egroups.com
This archive was generated by hypermail 2.1.5 : Fri Nov 01 2002 - 15:30:50 MST