AI Standards: Has the Time Come?

Claude Baudoin of cébé IT & Knowledge Management and Chair of the OMG's AI Task Force, chats with OMG Chairman & CEO Bill Hoffman about the advent of standards for Artificial Intelligence.

Visit the OMG Standards Development Organization at https://omg.org

Episode Transcript

(00:09):
Hello and welcome. I'm Karen Quatromoni,
Director of Public Relations forObject Management Group, OMG.
And welcome to our OMGPodcast series. At OMG,
we're known for driving industrystandards and building tech communities.
Today we're focusing on theObject Management Group standards
development organization,

(00:30):
and we're happy to behere with Bill Hoffman,
who is the OMG CEO and Chairman whowill lead today's podcast session.
Hi Claude. Can you brieflyintroduce yourself and your role?
Yeah, sure. First, thanks, bill forhaving me. My name is Claude Baudoin.
I am the owner and principal at asmall consulting company called cébé

(00:54):
IT and Knowledge Managementbased in San Rafael, California.
What we do is enterprise architecture,business process modeling,
managing the knowledge lifecycleand communities of practice for
clients. My background is in softwareengineering and IT management.
I have experience in the semiconductorenergy industries and I've

(01:17):
been associated with OMG for over 30 years
now. And I've been hangingaround OMG for years.
And now I am co-chairingthree groups within OMG,
the task force on businessmodeling and integration.
Also a working group calledthe Cloud Working Group,

(01:38):
which publishes guidelinesfor cloud users.
And very importantly for thetopic you have in mind today,
I'm also co-chair of the AIplatform task force for OMG.
Wow. And we want to talk about AI today.
So that's great and congratulations andthank you for being a member with us for
over 30 years. Very much appreciate that.

(02:01):
So AI has exploded over the last coupleyears. I mean, I wouldn't even say few.
I'd say even the last coupleof years it's just been crazy.
And it's obviously pretty much in flux,
whether it's the emergence of generativeAI concerns about trustworthiness and
bias. Does that mean it's tooearly to talk about AI standards?
Oh, that's an excellent question,

(02:22):
which for decades we've beenwondering when standards are
premature versus when they'retimely versus when they're overdue.
So I just would like to take a fewseconds and tell you an anecdote. First.
In 1993,
I was at an OMG meeting outside of Chicagoand there was a team of people I was

(02:42):
part of and we were writing acouple of booklets on comparative
analysis of the many object-orientedanalysis and design methods
that existed at the time.
And one of the methodologistswho had invented one such method
excoriated us in an articlein a well-known publication in

(03:04):
the object-oriented world.
And he wrote a piece called PrematureStandardization Considered Harmful,
which of course was a joke about goto considered harmful from the days of
structured programming. And twoyears later, just two years later,
in 1995,
that same person was sittingin a room with his colleagues

(03:26):
in San Jose,
California inventing theunified modeling language UML,
which was basically the standardfor object-oriented analysis and
design. So at which point is a standardpremature versus at which point is
it timely?
That anecdote shows thatsometimes it's a very short time

(03:47):
interval to the point where peoplerealize they need a standard.
So I think the fundamentalissue is this. Well,
there's a point at which peoplewho are applying or using a
certain technology arewasting a lot of time
manipulating data handling issues of

(04:07):
portability and interoperabilityand integration because there is no
standard and there is multipleproprietary formats and languages,
et cetera. And instead of focusingon the innovation they want to do,
they have to spend 80% of theirtime solving interoperability
challenges. That's when we need standards.

(04:29):
And there are some areas of AIthat are reaching that point,
which is why I think it'stimely and not premature.
Our job is to figure outwhere that border is and
how to not stifle innovationthrough standards,
but to allow people to focusmore of their energy on that

(04:52):
innovation by resolvingfor them through standards,
the portability andinteroperability issues.
That makes sense. I mean,standards provide leverage.
Leverage allows productivityincreases. That all makes good sense.
So can you give me examples of areaswhere you think these standards would help
AI developers and also the users?

(05:12):
So one thing we're working on right now,
we're just starting to work on itwith the help of one of our members,
a company in Austriacalled Zephyr Solutions.
We're looking at can people port models of
neural networks fromone platform to another?
When you look at convolutionalneural networks, CNN, today,

(05:36):
when you train a networkusing some training data,
generally speaking,
you must execute then the modelon the same platform with your
real data in order to getthe network to provide
recommendations.
So you cannot lift and shift the modelof the network from one platform to

(05:58):
another because there is no standardrepresentation of that model.
What do I mean byrepresentation of a model? Well,
a neural network has a certainarchitecture has a certain
topology. There's a number of layers,
there's a number of nodes on each layer.
Each node executes a certainactivation function on its inputs to

(06:23):
produce it outputs.These functions have parameters.
Can we create a model of allthis so that we can take a
model and move it fromone platform to another?
So that's the most exciting thingwe're looking into right now.
There's another area that we've beenlooking at, but we haven't yet found,

(06:44):
if you wish a leader to tellus what model we could develop
it's image classifiers.
So when you want to trainan AI to recognize images,
think of an
autonomous car company that wantsthe camera to recognize cars

(07:05):
versus pedestrians versus bikes.
You train the model onthousands of images.
These images come with a descriptionthat says, this image contains a car,
that image contains a pedestrian.There's no standard for this. Or rather,
there's a couple of different defactostandards that are not quite compatible
with each other.

(07:25):
So that's another area where we coulddevelop a standard that would allow
the developers of such modelsto ingest thousands or even
millions of images withouthaving to spend months
massaging the data so that
it can be ingested by the model.

(07:46):
And we have several other suggestionsthat have been made, for instance,
metadata to describe the datasets that are used to train models
in other areas than image classification.
So actually that goes withsemantic tagging of information,
which is a widespread need right now.

(08:06):
There are other efforts in other OMGgroups that have to do with standardizing
metadata to handle the semantics aspect of
this. So these are some ofthe things we want to work on.
It's going to be an amazing nextfew years, I can tell you that.
How would you determinewhich AI standards,
including some of the ones youjust talked about, might be needed?

(08:28):
So that's a good question becauseOMG works very differently from some
other standards organization.
Instead of having acommittee of OMG members
or employees of OMG, if you wish,
in the proverbial smoked field room,
deciding by themselveswhat standards are needed,

(08:51):
we have a very open processat OMG, as you know.
So of the things we can do isissue a request for information,
an RFI to find out what organizationsneed and that's going to drive
our priorities. And whenwe issue those RFIs,
they go to the general public,not just to the members of OMG,
so we can collect input fromthe entire world. In fact,

(09:15):
in 2019, five years ago, as we speak,
the National Institute forStandards and Technology,
NIST in the US had issued such an RFI
and they had 98 responsesincluding one from OMG.
The problem with that RFIis that it's five years old,
and in the meantime things happenincluding the explosion of large language

(09:39):
models, LLMs or generative ai.
So the landscape has changed.
And the other thing is NIST's missionis standards for the US government.
So the way some of the questionswere worded did not necessarily
address or incite responsesfrom an international audience.

(10:01):
And OMG is very mucha global organization.
So we have a process to gather more input.
And we're actually thinking of issuinga new RFI to answer your question,
which is to determine whichstandards might be needed.
Excellent.
So I know there's a lot ofother organizations that are looking at developing AI
standards. How would you thinkour OMG efforts differ from those?

(10:27):
When we speak specifically about ai,
there's a lot of stuff goingon right now, both at ISO,
the InternationalOrganization for Standards,
which has an entiresubcommittee devoted to ai.
It's called SC 42.
And then the IEE StandardsAssociation is also doing some work

(10:49):
on ai. So when you look at ISO,
they have a large numberof subgroups within SC 42,
which is working on various standards.
But when you look at ISO standards, alot of them are basically guidelines.
They tell people, thisis what you should do.

(11:11):
They're not very precisestandards if you wish.
And OMG works much more onformal standards that include
a UML model, maybe an OWL ontology,
maybe a specific languagedata formats, and
a graphical notation forthose models, et cetera.

(11:31):
So OMGs work is much moretechnically precise and
gives a lot more concretedirection to users and
developers on how they shoulddo things as opposed to general
advice on thou sh do XIEE standards

(11:51):
association. We have aliaison with them too.
I should have mentioned that wehave liaisons with many parts of
ISO. So we work together, we collaborate,
we feed a lot of OMG specificationsto ISO so that they become
international standards.So we are working together,
IEEE.

(12:12):
I am also co-chair of a project group on
AI terminology and data formats.It's going pretty slowly.
And again, it's going to be a glossary.
It's not a technicalmodel of an area of ai,
which is what OMG does.
Very good. Well,
you would think after 35 years wefigured out how to build standards,

(12:35):
and I think we do a good job at it.How can our listeners get involved?
Well, I mean, bill, thanksfor asking the question.
You would be probably even morequalified than I am to tell the
listeners that that OMG is amember driven organization.
So it's not

(12:57):
a couch potato opportunity.
It's an opportunity to get skin in thegame and get involved. So of course,
people can subscribe to theLinkedIn page or the mailing list
and read periodic updates.
But really people who understand theimportance of standards need to get
involved in creating those standards.

(13:18):
And the best way to do that is to becomean OMG member and to start attending
our quarterly meetings and themeetings we have in between
in order to generate theRFPs and then the standards.
So obviously the OMG websiteexplains all that to you.
You can find informationon how to become a member.

(13:40):
There is a cost associated with that,
and there is obviously a cost toattend meetings including travel. So
you need to start by understandingthe benefits to your organization
to justify those costs andthat membership. And to me,
it boils down to this,
do not wait for others to create astandard that might constrain what you do.

(14:05):
Then you suddenly discover thatmaybe your customers are asking you
whether you comply with that standardand you haven't seen it coming.
If instead you get a front seat andyou are part of the performance,
so to speak,
and you are part of the teamthat solicits and creates

(14:25):
and selects a standard, thenyou are shaping the solution.
And then you have advanced knowledgeof the way the standard is shaping
up.
And that'll give you anadvantage in the market once
the standard is adopted,
because you will havestarted developing your tools

(14:45):
or developing your practice as a user,
as a function of the standard yousee being developed because you're
participating in it. Soyou got to be an actor,
not a passive spectator.
Right. A leader. Not a follower. Thankyou, Claude. I appreciate your time.
You're welcome. Glad to have beenable to discuss this with you.

(15:09):
Thank you both. And so today you'vebeen listening to Claude Baudoin,
speaking about AI standards,has the time come.
And Claude is representing the ObjectManagement Group Standards Development
Organization. Thank you.