Re: [Heartlogic-dev] RE: FW: Research in emotional AI

[William L. Jarrold]

On Thu, 11 Mar 2004, Josh White wrote:

> > > preference is to think about the problem myself instead of
> delegating
> > > that task to a "smart" computer.
> >
> > I disagree.  But we can work on that issue in time....Ah
> >
> > I believe that we are collecting data at the website to maybe
> > help us give feedback to a learning algorithm.
>
> That's a good point and I agree it's a cool thing about your website.
>
> > it, or some combo of the two.  But what is an important aid
> > to doing either one of those is having a set of data that can
> > evaluate a given model (whether that model is created by a
> > human mind or in silico).
>
> With that in mind, maybe an Java API to access the data, along with a
> challenge to all comers a la Turing, would be a cool next step.

Josh, another excellent point.  Thanks!

I guess it is best to have a bunch of data for a given model to fit,
right?  Thus, we should not issue this challenge until we have a
sufficient amount of challenge data?

Then again, we can have a model receive a score based on its overall
believability rating.

In other words, there are two means of evaling a model:

1) fit the model to existing data

2) have the model *live* and have users rate its ouput on
inputs that are...(1) user submited or (2) randomly generated
or (3) other means of generation of inputs....To clarify, inputs are
scenario cues.

But, I like the idea of TURING CHALLENGE...That will catch people's
eyes and hopefully their brainpower....So, maybe it is not too late
to ad somethign to website in this regard.  Do either of you
feel like proposing a text blurb.

I was thinking of asking for the following top top level text blurb to be
added to the title page...This would be the first thing that anyway would
read...

"Our main goal is to help build AI's that understand human emotion
better.  In so doing, we will build something <cool>, something <useful>
and something which helps us deepen our understanding of ourselves.  We
hope to make progress in this direction by tackling what we believe is a
<small but tractable sliver> of the problem."

...each of those <blah>'s would be a hotlink to some text below.

<cool> I'm not sure what would be at the receiving end of this hotlink.

<useful> the receiving end of this would be a blob having several
sections.  One section of this blob would be the currently written
Clinical Utility section of the website.  Other sections of the blob
would include to be written portions about business relevance such as
games with realistic and sophisticated (organic and generative are great
marketingese descriptors) and training executives ($$$ !! (-;) to be more
emotionally intelligent ala Goleman (Mr All-The-Rage).  What happened to
that email I wrote about that. (Thanks Josh for suggesting this
incredibly basic incredibly important question!).  Whatever.

<small but tractable sliver> Would point to the stuff about hot vs cold
models.  That is already on the website, right Joshua?  It should say
if it doesnm't already...we believe that we value hot models, but for the
time being, we are mainly interested in cold models.  Focusing on cold
models is one way to take a small but tractable sliver of the
problem....Another way of taking a smallish sliver of the problem is to
frame the problem as scenario cue => appraisal.  And appraisal => valence
plus explanation.  Josh, not important for you to fully grok this part
for now...Joshua prolly gets this since he's read the dissertation...Now,
here is the new part: So, given this scenario cue => appraisal framework,
we can now control the complexity of the problem.  One crude control is to
limit to complexity of the scenario cue to e.g. less than 15 words such
that each word is a 5th grade reading level or below.......So, the
conclusion  of this blurbette (i.e. one of several sub-sections of the
<small but tractable sliver> might be, "So, our near term goal is to set
up a competitive framework for the challenge of modeling appraisal
processes that generate believable appraisals of scenario cues less the
15 words long s.t. each word is 5th grade level or below"...Farther out,
we wish to understand how the cost of this problem grows as the problem
space gets more difficulty.  E.g. how hard does it get to make gains
in believability with 7th grade vocab?  with 20 word scenario cues)...
We'd have to weave in a flashy blurb on the turing challenge in there
too.

Okay I have gone off too long now.  Joshua, assuming you find these
"friendly amendments" do you have the energy to clarify these ramblings
and get them up on the "front page" of the website? ... Then again,
maybe all that laid out is too ambitious, giving away too much of the
surprises in store?  Nah.

<...>

Bill