This is Paola from the AIKR
Community Group, Artificial
Intelligence Knowledge
presentation at WC. And it's
June. Finally, I get around to
show something. I've got
something to show finally.
So, about the slides very
quickly. To provide an
introduction because a few
people are new to this work and
they may still haven't quite
grasped what is being done and
why.
And myself, I also, as this
goalposts were shifting and the
ideas behind this work were
kind of shifting. I haven't
been sure what I was doing for
quite a while. So, it's just to
say this is what we're doing.
And also, there are a few
people who would like to
contribute. I thought we should
make it easier for them. Also,
there have been a couple of
questions that I hope this
slides will have to answer.
Requesting expression of
interest. This is very
important. So, people are
saying, do we have funding? Why
do we have funding? And because
I wasn't able to be very clear.
Now, what can start? It's
opening the proposal.
And I'm hoping there will be
people will be people will come
on board who are capable of
articulating budgeting and all
of the things and handle the
bureaucracy involved. Maybe
institutional partners or, you
know, backers of all sorts.
And also, it's going to be a
way of opening up feedback so
people can see what has been
done with a few slides and
finally be able to give
feedback on what has been done
and contribute.
So, people who have been in the
CG, the community group know
most of what has been going on.
You just search W3C AI KRSG,
you get to the landing page,
you get to the wiki. Mostly, it's
been discussions. But there has
been quite a lot of discussions.
And there's been quite a lot of
updates. So, the main task was
try to understand what is going
on in AI.
Because this has been more easy.
So, around the time when we
started is the time when
symbolic AI and knowledge
presentation were being nailed
in the coffin by the Turing
Prize scholars. And at that
time, myself, a few people
started kind of, you know,
getting nervous.
And they were saying, "where is
this going?" And soon after
that, AI, how do you say,
explainability risks and it all
became mayhem of where, you
know, what is it going to
happen and how we're going to
fix it. And a lot of people
have jumped on the AI bandwagon.
And WAGOR fits. Still not
understanding AI, which is
important. And I don't make any
claims that I understand AI,
but at least from this time
onwards, it is documented on a
mailing list that I've been
paying attention, at least to
whatever I could understand.
Out of the need to understand
machine learning, KR was
involved. So, when we couldn't
figure out what was happening
in machine learning, the black
box, we needed to call upon
knowledge representation to
figure. And at the time, people
still weren't sure outside the
niche of the KR people.
We were able to communicate. So,
we were able to communicate
what was. And out of the need
to understand machine learning,
KR was needed. But I don't
think anybody knew exactly how
it was going to be used to
achieve, to control the AI risk.
We still don't know, but we're
working on it.
This is that progress.
Took us a few years.
So what is KR?
We know that.
You can read the rehab risk.
Excellent education materials.
I don't have to tell you.
You say, okay, I'm going to
school.
I've been studying KR for years.
I've been a researcher.
I'm a professor in KR.
There are a few, not many, a
few people.
But as you can see from the
educational resources,
which we are creating on the wiki,
there are a lot of courses in
knowledge of recitation.
Some of them are excellent and
they're accessible.
And there are books.
So one can say, okay, I'm going
to study KR.
But that doesn't mean that
people who know KR, the
professor or even the experts,
know how to use it to address
the current AI concerns.
Or can't, you know.
So nobody can do anything in
absolute terms because AI is
evolving, how do you say, organically.
Nobody has control over, you
know, these algorithms are out
there, the computational power
is there.
And a lot of wonderful things
are happening, a lot of useful
applications.
So we know that we need KR for
accessibility and safety in
general.
So we're bridging the machine
learning.
So now that you see, KR
typically has been used to make
the machine to program machines.
So we wanted an intelligent
function.
We need KR to tell the computer
to execute it.
Now the opposite is happening.
We tell the computer is
executing an intelligent
function.
And we need KR to figure out
how it's doing it.
So I think this is very
exciting to see for me.
And this is what makes it worth
it, really.
So in order to communicate what
KR is, is such a huge domain,
we need this vertical because
there is so much of it.
And we need to also explain the
role of KR in explainability,
which becomes a bit circular.
We have to open the black box.
And we have to use KR to create
reliable, safe, ethical, unbiased,
and all of that.
So now this is the other very
exciting bit, using KR in the
creation of agents. Non-trivial
because the KR is a huge
subject. So you can point
people to KR, educational
resources, and there are books
and exercises and lectures they
can follow.
But will they, is anyone that,
I don't know anybody who can
guarantee a safe, ethical AI
system, whether they have
studied KR or not.
So, you know, but still we need
to be able to, to summarize
what KR is or which concepts
and tools and methods in KR can
be leveraged to guarantee AI
safety.
Because even experts in KR may
not be familiar with the entire
domain or with the state of the
art.
So typical expert people I've
spoken with, they, they were
experts 50 years ago when
things have changed.
AI has changed, computational,
computer science has changed.
Because using KR to make the
black box transparent may
require advancing the state of
the art in many fields.
KR but also logic. And this is
another exciting area. So, at
least we, so we have to start
by mapping the knowledge domain
for KR.
So ultimately we're doing an
ontology for knowledge
representation.
The scenario, important thing
to keep in mind is that AI,
whatever is going to be, which
there is a lot here, which
there is so much here I cannot
describe.
So, AI is actually driving and
influencing human evolution,
cognition, information, belief,
decision, the politics, the
economics.
So AI is not just by itself. It's
something which is very
fundamental in human
contemporary evolution and in
social technologies.
So, KR can help to mitigate
existential risks. To be honest
here, I have a whole talk on
this. Systemic aberration.
So, what is happening now is
that things are being done
which are not being described
properly.
And things that are said that
are not being thought. There is
a lot of confusion by mislabeling.
And we can only use knowledge
representation to help us disambiguate
and get a grip with what is
going on.
So, if you want, I can share
the link if you cannot click it.
If you're interested, that's
about an hour tall.
If the standards being
developed to ensure AI safety,
and I have sifted through a few
AI standards on the hub, are
missing out on this essential
AI safety concept, then the
risk of critical failures in AI
safety increase.
So, I hope this is fairly
logical and clear, despite a
lot of effort. So, there is a
lot of money and noise in AI
safety.
Nonetheless, the safety
standards are missing out on
essential safety concepts. And
therefore, I think we need to
use leverage KR to get out of
this loop.
So, I'm mapping the KR domain,
simply because when people ask
me what is KR, I want to point
to something that is cognitively
can be processed by a human and
by the machine obviously.
But, at the moment, KR is all
over the place. So, I have
divided into sub domain.
So, I'm saying the knowledge
presentation domain is made up
of upper foundational ontologies,
knowledge presentation
languages and formalisms,
domain ontologies. Oops,
spelling mistake here.
So, I can never remember what
ODDS stands for. Domain? Ontology
domain? Basically, these are
domain ontologies. And there is
a whole field and I'm looking
at it. When I can remember what
it is. AI safety standards. So,
here is where I found the
critical issues. Some critical
issues.
which are my worry, and I think
they should be your worry too.
So we have a lot happening in
the knowledge presentation
learning and ontology-driven
agents, very exciting stuff.
We cannot conceptualize all
this massive field without a
bit of sweat. So this is what
we're doing now, and I
appreciate very much your
attention and the concentration
you're applying to follow.
Now, this is the point where
there was a bit of a question.
Peter Rivett said, "Oh, you're
saying that you're going to do
something that can make the
system reliable."
He wasn't quite sure this is
something to be taken seriously
or feasible. So I thought, okay,
we cannot have any AI ethical
safety standards unless...
Unless we understand the notion
of reliability engineering. So
I'm intersecting these two here
and I'm happy to look at it
later.
So today we are starting from
this. We are modeling. So these
are subdomains. Each of these
is the subdomain for the
knowledge representation in
that domain, which is huge.
And today we are just focusing
on this state. Because I think
we should take this step by
step. So first question, are
the subdomains adequate? And
this is a nice little
evaluation task for anyone who
has a spare neuron.
So, okay. So, upper ontology,
knowledge representation
languages, domain ontology,
knowledge representation
learning, and all of these. So
is there anything in this subdomains
that shouldn't be there, that
doesn't belong? Or is there
anything fundamentally missing?
Please let me know.
And so what are we doing? What
I have done so far is that I've
taken all the upper top level
ontologies that I could put my
hands on.
I have identified the sources
of knowledge, which were
textbooks and papers and
repositories and abstracted
terms. In particular, I have
extracted the classes, I think.
And so now it's a matter of
evaluating. So then we're going
to develop a definition for
each class or category.
or streamline the duplicate
clean up boil it down a little
bit make a list and effect well
so when I
say list it looks like I'm
doing nothing you know but I'm
looking I'm making a bag of
words which is
a fundamental thing and I'll
pay mm-hmm so I have you can
click here if you cannot I'm
gonna send a
link separately a parentology
book up typically I would put
this in an open spreadsheet
open for
anyone to open but because
there have been people who have
kind of taken advantage of the
openness we
don't know who they are we have
there is no way to track whose
access to what resource maybe
using it
without credit to see doesn't
matter because ultimately these
are going to be published and
licensed for everyone to use
they're gonna be on an open
access but in the meantime some
people could
just use this work to do their
own thing and not acknowledge
no so this this is
unfortunately
happened so here is the upper
ontology vocab if you would
like to look at it or provide
feedback in
any way please request and if
you cannot see it please
request access I will however
already managed
to I will send the vocab to the
people who are already working
on this one so two or three
people have
already provided feedback on a
drums and these will receive it
probably in an email with a
request not
to share it outside at the
community group at this stage.
Please check the current list
of terms.
Annotate it. So in the Excel
sheet you may be able to create
a column with your name
and then in the column you
could write something for each
term, something like add,
you know, this doesn't,
whatever. So feel free to annotate
each term or to create an
annotation that
applies to the entire
vocabulary or even just email.
So are there any terms in cost?
About 200 terms now.
Anything to be added? Anything
that should be deleted? How can
these upper ontology
vocabulary, upper ontology be
used? Is it useful? Complete?
How can it be used? Make some
use cases?
Then eventually this is going
to be published and it's going
to be constituting the first
step into
the publication of a series of
vocabularies which overall will
represent the knowledge
representation
domain. And probably the basis
for an ontology of knowledge
representation. So if anyone
interested
in working on the list, please
get in touch. Thank you so much
for listening. And I'll see you
in the group.
Wow, this is being disclosed
already. Look at the machine.
This is the machine. It's doing
it.
The machine is opening. Bye.