[Aleader-dev] Re: dissertation feedback

[William L. Jarrold]

On Wed, 16 Jul 2003, Joshua N Pritikin wrote:

> On Tue, Jul 15, 2003 at 01:22:20PM -0500, William L. Jarrold wrote:
> > > I propose an emotion predictor that can be tested empirically.
> > > My primary interest is whether statistics suggest that it
> > > works empirically.
> >
> > Right.  That is where my dissertation might be especially helpful.  I
> > won't assert that doing the baby study I did was a huge big deal.  But,
> > what it did accomplish was establish a experimentally and
> > statistically and practically rigorous foundation on which to build many
> > more subtle studies.
>
> In a word, "interesting."  I have a bunch of comments and some
> questions:
>
> + 1.3 Methodological Issues -- Isn't this obvious?  I mean, the
> problem with psychology is that people can't figure out _how_
> to achieve procedual adequacy.  Surely researchers recognize the
> importance and benefit of it.  Or am I naive?  ;-)

Sorry, my politically incorrect impression is that most psychologists
either a) don't care about computers b) are afraid to learn them.  There
is a HUGE sociopolitical infrastructure that has established itself
and there are plenty of people to hire you, give you tenure, read your
papers even tho procedural adequacy is hardly mentioned.  It does hit
some people's anti-positivist and reductionist button I suppose.  But
this is a cynical reason and there are probably many non-cynical reasons.

Whatever.

>
> + With regard to theory of mind, the Aleader model carefully
> distinguishes between 1st person, 2nd person and 3rd person
> perspective.

Not sure I see the distinction.  Guess I need to read more about the
Aleader model.

> I equate "Theory of Mind mechanism" with a sense of
> empathy.

More or less yes.  A rational or "cold" form of empathy.  But it does not
nec invovle "hot cognition".  It does not nec involve feeling the other
person's despair in your gut.  This aspect of emotional modeling is
interesting but a much much harder problem.  Many smart people are working
on it and writing a lot about it.  It is too much like studying
consciousness for me.  I like smaller technical problems.   Maybe someday
my cold cognition type theory will help the hot cognition people do their
thing with more precision.

This is just my take on what TOMM is.  I don't know whether Simon
Baron-Cohen would agree with this or not.

> I appreciate your comments on autism.  While not my
> main interest, I also hope to contribute some deeper understanding
> to this clinical condition.

Sure.  I am in a Counseling PSychology PhD program.  There is quite a
large distance between clinical psych and ai.  Autism seems like on of the
few possible bridging areas in the nearer term.  In the longer term, the
bridging territory is huge.  But right now, the gap between the two
disciplines is painfully large.

>
> + The estimation of desirablility is also central to my emotion
> predictor.  I also sort desirability into three categories, which
> roughly correspond with your happy, indifferent and sad.
>

Cool.  I look forward to reading about this.  I think you sent me a
pointer to the readings.  If not, please do so.

> + Study 3 with the computer model is _incredible_.  That's exactly
> what I want to do.  Wow!

Great.  It is a nice feeling to have someone appreciate this.

>
> + In all three of your studies, you ask the believability of various
> things.  Believability is a subjective judgment. I try to explicitly
> confine and minimize subjective judgment.  Firstly, I start with the
> assumption that even "normal" people are frequently unable to assess
> situations accurately.  I want to use a computer model (like your
> Study #3) to propose _the_ precise method for assessing situations
> objectivity.  For Aleader, the subjective believability question
> comes in only _after_ the situations are categorized by emotion.

Hmm.  I believe that emotions are not right or wrong.  They are just
there, like the weather.  They are reactions that happen to you.  Now,
the feeling coupled with the explanation, that is where you get a right
and a wrong.  But for any given situation, there are a zillion possible
feelings and a zillion possible explanations for each of those feelings.

>
> + I am not very happy about explaining diversity in appraisals with
> the idea of generativity.  For Aleader, I have tried to eliminate
> diversity in appraisal by strict adherence to repeatability.  In
> practice, repeatability is achieved by appraising a recorded audio
> video experience -- film.

Hmm.  You mean you watch a film and try to figure out what emotion the
actor is experiencing?  Because the film doesn't change upon subsequent
viewings they experience the same emotion each time.

>
> + In your conclusions, you state that "mindreading" is considered
> a cognitive distortion, attributed to Beck (1995).  Can you briefly
> explain what Beck is talking about?  (Otherwise I'll look it up
> myself.)

Yes.  Imagine you and I are in a group therapy session.  You just said,
"I am not very happy about explaining diversity in appraisals with
the idea of generativity."...If I were paranoid, then I might say, "You
are so mean, why do you hate me so much!"  At that point you could
accuse me of dysfunctional mindreading.  Delineating the precise boundary
between dysfunctional mindreading and the sort of everyday mindreading
that is necessary for adequate social functioning seems pretty difficult.
Counseling and clinical psychology concepts are pretty mushy.  That is
what makes them so interesting to attempt to model.  If you can capture
them with appropriate richness, believability, what-have-you, you can
capture anything!  Sometimes I believe the goal is to replicate (A) the
fuzzy-headed thinking of everyday humans.  Othertimes, I believe the goal
is to do (B) "better" than that.  However, an AI that could do a
half-assed job of (A) would be much more interesting than an AI that could
do (B).  Sorry, I am starting to digress.

>
> + What is your opinion of Knowledge Machine (KM)?

I like that it has Self.  This is like Java's "this".

I like that it allows one to unify objects.  For example, thus, if you
assert "every animal has a head." and "every dog has a dog head" and
"a dog is a kind of animal" then when you instantiate a dog,
it will unify the skolemized dog head with the skolemized animal head.
Cyc will skolemize two heads - an animal head and dog head.  This
can be overcome, with some more assertions but I am not sure if those
assertions are currently supported in Cyc.

Bruce Porter and Peter Clark the main creators of it think harder
and more principled than the creators of Cyc, I think.  Their user manual
is a joy to read.  The most fun I have ever had reading a user manual.
Not only does it teach one the nuts and bolts of using KM, it also raises
some great KR issues nicely and succinctly.

One example of this level of principledness is that they endeavor to have
code support for every feature in the KM language.  Cyc, on the other hand
has taken the following approach "lets make assertions first and worry
about code support later.  So what if we can't test it now."  They also do
not do a good job of documenting which parts of the ontology work and do
not work.  E.g. #$holdsIn -- it expresses temporal bounds on propositions
-- is one of the oldest and most important predicates in the system and
has never been code supported.  Internal Cyc has recently (within the last
year) begun supporting temporal reasoning (using different predicates than
#$holdsIn).  This is a very exciting development.

KM It is a much simpler system.  Some might say too simple.  The
unificationcan be a bit heavy handed.  It sometimes smashes two objects
together (i.e. unifies them) when they should not have been smashed
together.

It is also frame based.  This can cause problems.

Sorry I am being brief and handwavy because 1) I need to work on things
which pay me money 2) It will take me a while to reconstruct this stuff.

> How does it compare
> to OpenCyc?

The gui in Cyc is much much better.  However, if you are an
experience computer person, you might pretty being "close to the metal" --
command line interface and all that.  If you like Windows, then you'll
like cyc's interface better.  If you prefer unix/linux command line
type interfaces then the gui won't be the deciding factor.

I have seen demo's of a very nicely designed GUI for the KM system.  It
was used in the RKF project.

Cyc is a much more mature system.  It has been pounded on for years by
all sorts of people.

Open Cyc has a teency weency amount of knowledge.  I checked the
other day and, c'mon #$Fruit and #$BillClinton are not even in it!
So many rules are missing from open cyc.   The full Cyc is much more
interesting.   There is talk of "Research Cyc".  If a credible proposal is
made that does not activate Doug Lenat's flake detector or his fear of
losing intellectual property, then he might allow one access to such a
gem.  Having worked first at MCC and then Cycorp when it started, I know
Doug pretty well.  I've worked for him for over 10 years.

> I see you worked for Cycorp as an OE for some time.
> Since I haven't written any CycL yet, I am certainly not attached
> to a particular inference engine.  Which software is most suitable
> for the kind of knowledge representation we are doing?

It is too early to decide and not important to decide now.  If you write
100 rules and then change your mind, you won't loose much time at all.  if
you write 1000 rules and then change you mind, well, then you might have
been better thinking more carefully first.

Cyc might be better to start with.  There's a bigger community of users ==
more help for KR issues.  Alas virtually no one discusses real KR issues
on Open Cyc.  Plus, having an interface at the beginning might even
be good for an experienced person like yourself.

Intellectual property issues also should play a role in the decision.
OpenCyc is, if memory serves, lesser gpl (whatever that is) and I don't
know what KM is like.

Bear in mind that I am a partial owner of Cycorp but I strive to look at
things fairly and objectively.

Can you tell me more about your prior background?  What kinds of coding
have you done?  What sort of I-bank were you at?  Where do you want to go
with this stuff?

Maybe a tel conv would be better?  Feel free to phone me 512-567-0738
after 9am or before midnight Texas (Central Standard Time).  If you have
a resme or cv, feel free to send it.  I have to score a neuropsych test
today and its my birthday so I am going out later.  But hopefully I will
resurrect my archived files and be able to send you some of the papers I
have written.

Bill

>
> --
> Victory to the Divine Mother!!         after all,
>   http://sahajayoga.org                  http://why-compete.org
>