Re: [Heartlogic-dev] do you have the reversal blues?

[William L. Jarrold]

On Thu, 16 Jun 2005, Joshua N Pritikin wrote:

> On Wed, 2005-06-15 at 10:48 -0500, William L. Jarrold wrote:
> > I've been thinking obsessively about reversal.  Maybe we should not have
> > a reversal condition.  Or at least not do it so much.  Well, I have not
> > convinced myself enough to drop it (or at least lower it) but thought
> > I'd letya know.
>
> I'd be more interested in knowing _why_ you have the reversal blues.

Lotsa reasons.  It forces us to ask a buncha really dumb questions.

I think we should keep marching.  I have written a long ramble on 
reversal.  It would be good to get the grouup thinking about it.
Forcing myself to talk about forces me to think about it more deeply
and grok it more thorugohly.

Maybe they will like the experience.  Maybe they will vomit so profusely
that it will be clear we should minimize reverssal.  We will only know if 
we we try.

> My
> guess is that it is hard to know how to reverse items in general.

Yes.  There are two main techniques at least...Techniques of
ablation, as in computational ablation. (It is better to call it
ablation than reversal, maybe)

One negate it.  E.g. "Jesus does not love his father."

Two put in a substitute/opposite e.g.

"Tracy wants an apple.  She gets one.  She feels sad bc she got what she 
wanted."

or

UNABLATED:

"If water touches a objecd tthen the object is wet."

ABLATED/SUBSTITUTED

"If water touches a liquid tthen the object is wet."


> So
> even if we collect data on reversed items, it is not clear what the data
> means.

If unreversed more believable than reversed then our hypothesis that the 
assertion or rule or whatever is correct is supported.  If not, then thhat 
is when we don't know what to do....E.g. maybe both p and not p are true
but it depends on implicit context.

> Is the item really reversed or not?

I don't understand the question.  Of course we know the item is reversed 
bc WE REVESED IT!!.

anyway, we really should not get into a philosphical debate when i need to
give a polished talk in less than 24 hours.

thus, lets keep marching on the reversal bandwagon.   okay?

bill

p.s.  i have included my reversal brainstorm rammble below.


REVERSAL

;;;;;;;;;;;;;;;;

* has roots in computational ablation.

(*) important in some cases to ask both p and not p.  a means of
    uncovering assertions that are ambiguous in NL.

(1) helpful if objects are poor. "if your robot is really
    unbelievable, you can at least prove that it is doing something by
    whacking the hell out of it."

To understand computational ablation, the first step is to understand
its origins in neurosience.  Neuroscientists like to do controlled
studies where the damage parts of an animals brains.  For exampoe,
they might hypothesize that area V1 in the rat's brain is critical for
stereo vision.  So they will take a knife and carefully damage area V1
for one group of rats.  For another group they will damage some other
area, and have an undamaged control group.  If their hypothesis is
correct then the rats that had damaged V1 areas will do
signifincatly worse on a vision task.

Now, in computational ablation, the idea is to do an ablation study on
some kind of computational entity.  So, think of taking poor Marvin
the robot and whacking the hell out of some part of his program.
You'd predict that he'd behave worse.

Of course, this is a dumb thing to do when your robot or your computer
program is very simple.  But when it starts getting very complicated
it starts behaving in unpredictable ways.  Bugs in ones program are an
example of such unpredictable behavior, but this is a side issue.
Think about it, when our computer program gets very "smart" or very
sophisticated we really loose our ability to grasp what is going on in
any particular piece of code....So, one thing we can do to partially
improve our understanding of this system is mash some part of the
program and see what happens...So, that is really what computational
ablation is.

Now how does it apply here?  Well, the object of our study is really a
theory of common sense reasoning.  Thus, if we hypothesize that ...

Foo implies Bar

...is some rule in our collective humann common sense knowledge base,
then we can subject this to a falsifiable empirical test.  If our
theory is correct, then:

"Foo and NOT Bar" cannot hold

...yet...

"Foo and Bar" can hold.

;;;;;;;;;;;;;;;;

Now, so that justifies putting in a not in front of various props.

But there are other kinds of ablation we might want to do.  For
example, in my dissertation, I wanted to established that emotion was
in some sense partialy irule governed.  So, insted of negating a given
conclusion, I put in an opposite conclusion.  E.g. If our theory of
emotion would conclud that Sally would feel happy we would ask our
subjects in the ablated condition if Sally would feel sad.

<show slide>

;;;;;;;;;;;;;;;;

reversal is most important for weak theory domains. or open textured
domains....in such domains truth is more graded.  also, you are more
likely to find cases where p and not be are both considered
believable.

also good for when behavior is arbitrarily bad.  unreverseed items
should suck less than reversed items but both unrev and rev may be 
UNBELIEVABLE.......it is good to get this kind of validation when our AI 
system might be especially impaired.

also good when the subject is controversial.  e.g. "jesus is god's son."

;;;;;;;;;;;;;;;;

And yet there are other reasons to look at reversal.  For example, you
might think that "Birds of a feather flock togher" or "Likes attracted
likes."  YET, "Oppsoites attract,"....Or "Too many chefs spoil the
batter." yet, "Many hands make light work." 
>