LAURENCE MORONEY: Hi, everybody.

And welcome to "TensorFlow Meets."

I'm Laurence Moroney.

And I'm delighted today to meet with Arun Subramaniyan

from BHGE.

And Arun, I know you've been doing lots of great stuff

with probabilistic modeling.

So for those of us who don't really

understand probabilistic modeling,

could you tell us all about it?

ARUN SUBRAMANIYAN: First of all, good morning.

And thank you for having me here.

And so probabilistic modeling and probabilistic theory

generally is something that we've been

using for several years now.

And it's mostly for modeling systems

that have a combination of very complex phenomena coupled

with things that we can't measure precisely.

And just to give you a simple example,

if I were to ask you to predict where the stone would land

if you throw it, then any high school student

would tell you that they can calculate it precisely

based on how fast you threw it and at what angle

you threw the stone.

Now, if I were to add a little bit of uncertainty

to it, saying I don't know exactly at what angle

you threw the stone at or what velocity you threw it at,

then--

LAURENCE MORONEY: Maybe like wind shear and stuff like that?

ARUN SUBRAMANIYAN: --and wind shear and stuff

like that, then all of a sudden, your predictions

are no longer as precise.

Now, in a simple system like that,

you can already see things starting to get complex.

Imagine a complicated system in the real world.

Things can get much more complex if you're trying

to predict something precisely.

LAURENCE MORONEY: So you're working

on a lot of complex systems like this.

Could you share some examples?

ARUN SUBRAMANIYAN: Absolutely.

So at BHGE and broader GE, we work

on a lot of complex systems.

So for example, say, designing gas turbines or trying

to predict what a very large-scale system like

an offshore oil platform would do--

we're talking about hundreds, if not

thousands, of variables interacting with each other.

And most of the time, you can predict

or you can measure maybe a few hundreds, if not

a few tens of those variables.

So how do you predict behavior of such complex systems

and still have actionable, meaningful outcomes

from your models without knowing all the information

about the system?

That's really where we use probabilistic modeling.

LAURENCE MORONEY: I see.

It sounds complex.

So what is your approach to this?

How do you get started?

ARUN SUBRAMANIYAN: [INAUDIBLE] So we get started with--

starting with the domain.

So we understand the domain.

So what I mean by "domain" is you

might be a mechanical engineer.

You might be an aerospace engineer.

You might be a petrophysics engineer.

You start with the understanding of the domain

and marry that with traditional machine learning techniques.

And that's been going on for several decades.

That gives you a very good understanding

of how to predict things precisely.

And that's what we call known knowns.

And so we can predict the known things very precisely.

LAURENCE MORONEY: So known knowns, right?

ARUN SUBRAMANIYAN: Exactly, known knowns.

LAURENCE MORONEY: A core area you can work from.

ARUN SUBRAMANIYAN: Core area we can start from.

And then we add a layer of probabilistics on top of it

to say, what are the things that we cannot measure precisely

or measure at all?

And that's where probabilistic modeling comes in.

And that is what I would call known unknowns.

An example of that would be, say,

if I am trying to predict how a crack is going to propagate

in a particular component.

Then I need to know what is the temperature

of that particular component, for example.

I measure it to within plus or minus 10 degrees.

But I don't know what that variation in temperature

is going to do to my crack propagating.

LAURENCE MORONEY: I see.

ARUN SUBRAMANIYAN: So that work, that's

what I would call known unknowns.

And once I know what are the unknowns that I'm not

entirely sure about, I can go say, OK, this

is what is the impact of that on something [? real. ?]

There is another level of complexity, where things

that I don't know I don't know.

LAURENCE MORONEY: Got it.

ARUN SUBRAMANIYAN: And that's what

I would call unknown unknowns.

LAURENCE MORONEY: Unknown unknowns, right?

ARUN SUBRAMANIYAN: I know it's a mouthful.

But an example of that would be, say, you

have designed a system.

You have put it out in the real world.

You know some of the things that is going to affect that system.

But you're not entirely sure of everything that's

going to affect the system.

And that is that other everything.

That is what we would call unknown unknowns.

LAURENCE MORONEY: I see.

ARUN SUBRAMANIYAN: And most of the time, in real world,

you can predict something up to, say, 90% or 95% of the time.

The last 5% is what surprises us.

And in systems which are safety-critical systems that

are critical to keep up the infrastructure of the world,

you can't necessarily have even a 1% chance

of something going down.

So for example, if power goes down,

you need to be able to bring that back up very quickly.

So those are the kinds of things where unknown unknowns come in.

LAURENCE MORONEY: Got it.

So starting from the known knowns,

then going to the known or knowable unknowns, and then

there's the unknown--

ARUN SUBRAMANIYAN: Unknowns.

LAURENCE MORONEY: Unknown unknowns.

ARUN SUBRAMANIYAN: Exactly.

LAURENCE MORONEY: I see.

So you've gone from known knowns to knowable unknowns, and then

unknown unknowns.

ARUN SUBRAMANIYAN: Unknown unknowns, right?

And when you're trying to model systems that are highly complex

and are extremely critical, you need

to be able to predict things at all of those levels.

And even if you're not able to predict unknown unknowns,

you need to know how much are you missing.

If an event like that happens, how would you respond to it?

That is really where unknown unknown comes.

LAURENCE MORONEY: So now, bringing this, then,

into just developing these things,

you use TensorFlow Probability.

ARUN SUBRAMANIYAN: Yes.

And we started with TensorFlow and then combined that

with TensorFlow Probability quite a bit.

LAURENCE MORONEY: So could you tell us a little bit

about how you use all that?

ARUN SUBRAMANIYAN: Absolutely.

So we started with TensorFlow for deep learning, precisely.

And when we got introduced to the TensorFlow Probability

team, and Josh Dillon specifically, what we realized

was they were bringing extremely deep research concepts

from the probabilistics world into a production world

that is not generally common.

And we were able to mix the deep learning community

with the probabilistics community we

had within our own teams.

So running a reasonably large data science

team, what you have to do is mix teams that are not necessarily

talking the same language.

And TensorFlow allows us to do that

very effectively because now, a deep learning expert who

doesn't understand probabilistics well

can talk to a probabilistics expert who

doesn't understand deep learning in the same language.

LAURENCE MORONEY: Nice.

So having that framework that they could work together

was very powerful for them.

ARUN SUBRAMANIYAN: Absolutely.

And it helped scale both our teams

as well as our deployments very quickly.

LAURENCE MORONEY: Wow, so a lot of complex

stuff that you've been working on.

There must be somehow you got started to figure out this.

And how did you learn all this?

ARUN SUBRAMANIYAN: Absolutely.

So I'm not a trained data scientist by training.

So I got into data science by accident.

So I'm an aerospace engineer who had

to solve very complex problems by mixing

these concepts together.

LAURENCE MORONEY: It's a very common story, by the way.

ARUN SUBRAMANIYAN: Absolutely.

So one of the things that helped me a lot

was trying to mix practical aspects with the deeply

theoretical aspects.

So for practical aspects, a book that I really love

is called "Doing Bayesian Data Analysis."

And that gives-- at least it gave me--

quite a bit of understanding of how these

are applied in the real world.

But at the same time, thinking about probability requires

people to think about solving problems

in a very, very fundamentally different way

because we are good at being trained at saying,

here are the bunch of inputs.

How is that going to get me one outcome?

But if the same inputs give you multiple outcomes,

that's a very different paradigm to think about.

So a set of books that helped me were

from E.T. Jaynes, which is at a much more philosophical level

of understanding probabilistics.

I would urge folks to at least dabble

in both the practical aspects as well as

some bit of the philosophical aspects together.

And if you look at the recent blogs from the TensorFlow team

as well as the broader community in doing

probabilistic deep learning, there's

a lot of fantastic blogs out there

that'll help people get started as well.

LAURENCE MORONEY: And I'd say one thing.

I know you've written a couple of blogs yourself.

There's another one on the way the way

where you've gone into a little bit more detail

than what you've been talking about today.

ARUN SUBRAMANIYAN: Absolutely.

And we walked through the three blogs

because we wanted to walk through known knowns, known

unknowns, and unknown unknowns, and bring it all together.

LAURENCE MORONEY: So the one that you're still working on

is the unknown unknowns.

ARUN SUBRAMANIYAN: Yes, and we're

close to getting done with it.

And it's getting published in the next month or so.

LAURENCE MORONEY: So all that's on blog.tensorflow.org, right?

ARUN SUBRAMANIYAN: Absolutely.

LAURENCE MORONEY: So thanks so much, Arun.

And thanks, everybody, for watching this episode

of "TensorFlow Meets."

If you've any questions for me, or if you've

any questions for Arun, just please leave them

in the comments below.

And also, in the description for this video,

we'll put links to everything that we spoke about today.

So you can check him out for yourself.

Thanks, and see you next time.

ARUN SUBRAMANIYAN: Absolutely.

Thank you, and thanks for having me.

[MUSIC PLAYING]