[Heartlogic-dev] Re: reactions

[Joshua N Pritikin]

On Fri, 2005-06-03 at 19:05 -0500, William L. Jarrold wrote:
> However, I am still 
> unclear on the basic research question...
> 
> E.g. at http://wiki.nirmalvihar.info/index.cgi?JoshuaGoalResearchQuestion
> 
> ...one would think that this page given its title would clearly lay out 
> the research question.  I see lots of interesting and helpful text giving 
> me background.  But the only question I see is "Which emotions can 
> conceivably be experienced by an ape and which are uniquely human 
> emotions?"...But surely a web based study is not the way to ansewr such a 
> question.  Rather, we'd need to play with apes to answer such a question.

Indeed?  I feel like I've answered this objection from you already.

What I thought I said is that Tomasello has already spent a lot of time
playing with apes, reviewing other people' research about apes, writing
a 500+ page book about apes, and being director of the Max Planck
Institute for Evolutionary Anthropology in the Department of
Developmental and Comparative Psychology.

http://email.eva.mpg.de/~tomas/

And what Tomasello has concluded is very simple: The only significant
difference between humans and other primates is joint attention.  And
even if he's wrong, joint attention is still an interesting phenomenon
to investigate (as I acknowledge in this email below).

So if the only thing I need to worry about is joint attention then I
_can_ do a web based study.

> Maybe the gripe in the above paragraph answered by labeling it not 
> researdh question but rathher "motivation" or "background and motivation"
> or "introduction."
> 
> Of course, you mention these as the research goals...
> 
> Overall Goal:
> 
> 1. To empirically induce a taxonomy of goal-pairs.
> 
> 2. To isolate an agent's proper goals from an agent's concern-for-others 
> goals.
> 
> 3. To gather believable examples of each goal-pair.
> 
> 4. To see which stories 'win' the competition for the most believable in 
> each goal-pair category.
> 
> 
> ...that's better.  But still I have question about this.
> 
> E.g.
> 
> For 1.  I thought we already had a taxonomy of goal pairs.  It is really 
> just a set of goal pairs and it has 9 elements, goal/goal; goal/anti-goal;
> goal/no-goal; etc etc.....So what is 1 a question?  Why do we need to
> empirically induce what we already have?

The 9 categories are just the beginning.  To me, what you are saying
here is similar to holding up the first chapter of OCC and asking why we
need to read the remaining chapters.  ;-)

> Hrm, maybe this is what you mean....Maybe you mean,

Yes, I mean "let's go deeper."  I don't have a specific plan at this
point but I have a specific direction (joint attention focused on
goals).

>  let's try hard to come 
> up with a no-goal/no-goal story.  Is such a story believable?...Maybe 
> weird existentialist stories like "Waiting for Godot" or "Rosencrantz and 
> Guildenstern are Dead" are examples of Goal No Goal Stories.

Yes, maybe.  I want to find out ...

> For 2. What is an agent's proper goals?

I'm not sure whether #2 is a research question.  I am planning to expand
on #2 by specifying as precisely as possible what I mean by an agent's
proper goals and an agent's concerns-of-others goals.  I'll probably
cite Elliott and use his terminology along with the terminology from
your dissertation and also some other ideas I have from grammar and from
Tomasello.  The point being to propose precisely how agents hold goals.

> For 3. It seems we really need to get to work and come up with the stories 
> ourselves.  I don't think this is something the average jamoke on the WWW
> can come up with.  No one will spend the time to read all our wiki's to
> figure it out.

No problem.  I have a big collection of stories, enough for each
category.  I invite story submissions from the public to help insure
that I'm not getting stuck with tunnel vision.

> For 4. I don't understand why this is a scientific issue.  I mean it would 
> be fun to know which ones are most believable in each category.  It would 
> also help us figure out the best examples to use when explaining our 
> taxonomy....But what would really help solidify the status of the 
> taxonomy is many good examples of each category.  And also clarity that 
> one exemplar does not go somewhere else.

It seems like you answered your own question.

The other reason for #4 is that once we have a collection of believable
stories for each category then we can think about refining the taxonomy
with additional branches.  For example, if both apple and orange fit
into category A then we can name category A as "citrus" and further
branch on the color of the fruit.  Something like that.

> Also, for 4, why not start simple and follow occam's razor.  If so, we 
> should first identify stories that win as the most believable for the
> unary status of goal.  Or the unary status of no-goal.  Or the unary 
> status of anti-goal.  It would seem to me that we have to do that before 
> we do goal pairs.

Your proposal is ironic because that is essentially what the current web
site is doing -- goal-pairs are created by combining two unary goals
using the min function.  Min usually works OK, but sometimes gives false
positives.  Eventually, I probably want to collect data on proper
goal-pairs instead of only synthesizing them from unary goals.

> Well, I'm curious to see how you react to my questions.  Maybe my ego 
> is getting in the way and I can't let someone else "do their own thing."
> or maybe my comments are construcitve.  I dunno.

At least recently, I feel like your comments are entirely constructive.
From the email, what should we add to the wiki?

> I generally feel that there IS something here.  There is something 
> important about these goal pair states.  They seem like components
> of affective states of high level organisms, like humans and perhaps
> some great apes and/or whales/dolphins.

Good .. even if joint attention is not uniquely human, it is certainly a
perspective that is accessible only to a high level organism.  Perhaps I
should justify goal-pairs more humbly, but I do enjoy the drama of
talking about HUMAN emotions.  I think that trying to distinguish human
emotions from other primate emotions is something that is catchy and
will help to attract participants to the study.  Also, as I mentioned
above, Tomasello has collected quite a lot of empirical evidence in
support of joint attention being a uniquely human invention.

> If we want to establish their validity as conceptual clusters that are in 
> peoples heads, one idea is that we could ask people to take a bunch of 
> stories and cluster them into related groups.  If people clustered them
> into 9 clumps that corresponded to the 9 goal pairs, that would support
> the hypothesis that these goal pairs do underly some phenomenologically
> important states...Does that seem like a core research question to you?

Sure, that sounds like a core research question.

However, I doubt that such a study could be carried out because of the
requisite rigor involved in evaluating goal-pairs in a story.  In your
previous email, you wrote out some examples of stories and I found that
you had set things up in a way which was not entirely consistent.  If
you can't do it helped by your familiarity with goal-pairs then I
strongly doubt that the public can.

> Trouble is that is a psychology quesiton.  not really an AI question.
> I'm not steadfastlly against doing such a study but I'm more interested
> in focusing on AI.

I assume you referring only to the study you mentioned directly above.

With respect to AI, I quote from
http://wiki.nirmalvihar.info/index.cgi?OpJoGoalDefinition

By confining the goal pair model to what is possible to recognize
physically, the possibility of using this model to help bootstrap
language learning is left open.  A robot similar to the one built by
Scassellati might be extended to understand goal pairs following the
methodology described in Vogt (2000).

-- 
If you are an American then support http://fairtax.org
 (Permanently replace 50,000+ pages of tax law with about 200 pages.)

signature.asc
Description: This is a digitally signed message part